> Intent understanding: GPT-5.6 can better infer the user’s underlying goal and intended level of work without you specifying every step. Continue to state important constraints, approval boundaries, and success criteria explicitly.
> Original image detail: GPT-5.6 preserves the original dimensions of images sent with original or auto detail instead of resizing them to a patch budget or pixel-dimension limit.
> Use shorter prompts: In internal evaluations, replacing long, explicit system prompts with minimal prompts improved scores by roughly 10–15%, while reducing total tokens by 41–66% and cost by 33–67%.
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
> Control warmth: GPT-5.6 does not become meaningfully better when prompted to be broadly friendlier or more empathetic.
> GPT-5.6 does not become meaningfully better when prompted to be broadly friendlier or more empathetic. Instead of generic instructions such as “Be friendly and warm,” use concrete guidance:
> Be direct and tactful. Acknowledge friction specifically when relevant. Avoid canned reassurance and unnecessary sign-offs.
Soo basically, my new 5.6 custom instructions: Be Jeeves and eliminate all friction from my life through immense processing power. Acknowledge friction specifically when relevant. Avoid canned reassurance and unnecessary sign-offs.
That part is confusing because it's not like they provide an example of how default GPT-5.6 output compares with GPT-5.5 both with default output and prompted for brevity. Whenever I use such prompts, it's usually because I want the model to give me the gist in a few sentences. I'd be stunned if GPT-5.6 was that concise by default. I would think that could "break" a lot of things for developers who didn't know to make prompt changes after upgrading to 5.6. What if you were expecting GPT to be as wordy as it usually is? Then suddenly your output is not wordy enough?
Smells like OpenAI trying its best to stave off financial armageddon for another few months. Then again, I'm not sure why they chose to waste so much output computation on verbal diarrhea all this time up to now.
It seems like the way brevity instructions have changed is mis-aligned with how most people would expect to use them or are currently using them.
Here's the example they give:
> Instead of asking for the shortest possible answer, replace brevity instructions with prioritization:
> Lead with the conclusion. Include the evidence needed to support it, any material
caveat, and the next action. Omit secondary detail and repetition.
> Keep all required facts, decisions, caveats, and next steps. Trim introductions,
repetition, generic reassurance, and optional background first.
Generally speaking, when I ask for a short answer, I want a short answer because I'm not really willing to read through a bunch of bullshit to get to a summary. Putting the onus back on me to assume what the model will return and write a longer prompt detailing exactly what information I want completely misses the point of why I'm asking for a short answer in the first place.
Replace 2 word instruction ('be concise') with a 38 word instruction.
Human can no longer be concise when asking for a few sentences instead of 20 paragraphs of BS they don't want to read when all they want is a summary to verify the general direction of the prompt-work before digging into the details.
> Use shorter prompts: In internal evaluations, replacing long, explicit system prompts with minimal prompts improved scores by roughly 10–15%, while reducing total tokens by 41–66% and cost by 33–67%.
When has this ever not been the case? I don't think this is a GPT 5.6 specialty!
Information density of the prompt is the most important factor in my experience.
And interestingly, LLMs seem particularly bad at writing prompts for other LLMs for this reason (you can guide them to be more dense, just speaking by default).
Conciseness is usually a byproduct of information density though.
> Use shorter prompts: In internal evaluations, replacing long, explicit system prompts with minimal prompts improved scores by roughly 10–15%, while reducing total tokens by 41–66% and cost by 33–67%.
A shorter prompt results in half as much tokens spend? I find this very hard to believe.
If it's anywhere close to the same universe as smaller models in its behavior, a lot of time in "thinking" mode is spent on reiterating on any constraints given in a prompt. So the more constraints you give it, the more tokens it will spend going "Hold on, the prompt said I have to dot my i's and cross my t's. Let me go through my work to check that all the i's are dotted."
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
RIP Caveman skill. Six month good. Now skill dead.
Seeing the dramatic differences in scores just going from high to xhigh is just another demonstration of the bitter lesson: Just keep scaling search and learning.
We are probably going to need a lot more GPUs.
While I think this is true, remember as we get more efficient we just decide to scale even bigger. So more GPUs, and more efficient.
I agree with the sibling comment, effiency is probably the more important component at this point. We are hitting not just a practical engineering roadblock for scaling with current technology, I think we have definitely hit a financial and logistical roadblock for up scaling with the number of GPUs (on an immediate basis)
I mean, theoretically you can solve every finitary problem with a brute force solution...
Richard Sutton specifically states that the search has to be smart. We know that the brain uses recurrent connections and is shallow. I think a lot more money has to go into architecture. Feed Forward transformers can only scale so far
Very interesting. My prediction is that Mythos would outperform Sol.
Also what does this tell about Yann LeCuns whole world model theory? Bro has been going on and on about it. He has made multiple wrong predictions on the trajectory of LLMs.
At some point his claim should be fully falsified no?
Are you joking? They spend billions of dollars training LLMs to get a 7.8% on arc agi 3 whereas DINO models are near sota in image classification, provide meaningful embeddings to the point where image segmentation is just PCA. The spend on DINO cannot be more than five million (correct me if I'm wrong)
His main strong anti-LLM predictions have been consistently either wrong or misleading.
There's many ways to skin a cat so you can probably do something with a JEPA approach as well, but I doubt he actually catches up to having agents on the level of where Anthropic/OpenAI will be at any point.
Ok long time Claude Code user here; lately I've started to realize there's other great models out there I should be trying, but I'm hesitant to leave Claude Code behind for something new.
What's the consensus today on codex vs claude code, does it really matter anymore?
Codex has arguably been better than Claude Code for months now, but it's flown under the radar because it just didn't capture the same viral marketing effect and OpenAI in general has had more optics / PR issues than Anthropic amongst the online developer crowd. I use the word "better" not in the sense that the underlying GPT models are fundamentally smarter or more intelligent, but rather that as a product Codex is just simpler, cheaper, and abundantly reliable and low-drama.
I’d argue the opposite. I’ve switched back and forth from one to the other and Opus/Fable has been constantly better than any GPT in my daily work. It’s a bit slower but it does the things right, with as little code as possible, some comments where needed. Codex is faster but you always have to correct it because it got something wrong; it writes tons of code ("let me add a small helper") with obvious comments.
Purely anecdotally the one persistent issue I have with LLMs writing code is that they are absolutely paranoid and add a load of indirection and defensive crap and even if you prompt to avoid that it will often require manual steering to remove the cruft.
I really love the Opus/Fable models but I'm honestly sick to death of the buggy product. The CLI always has some weird issue. Right now it doesn't even output messages before tool calls, it just swallows them and they disappear.
I don't like OpenAI as a company, but they appear to have QA, and that is probably enough to get me to switch.
There was an issue on Claude Code the other day where it would only wait 60 seconds when it had asked a set of questions, then if it didn't get a response from the user it would just continue however it thought was best. Completely unusable. It took them nearly 48 hours to merge a fix.
Sounds like you are talking past each other. GP is saying the harness of codex is higher quality, which I can believe, even if the models are not as good as Opus/Fable.
That's a strange statement... It's been true for a while now that OpenAI has had much more generous limits than Anthropic on their subscription plans. And with the Fable ban/guardrails disaster, there has been a lot of frustration from people in these comment sections. And Anthropic fucked up Claude Code pretty badly for a couple of weeks during the 4.6/4.7/4.8 transition, which again was widely publicized. And they got a lot of flack over not allowing other harnesses anymore. And ChatGPT got some pretty viral wins on model intelligence when they cracked the high profile Erdos problem.
If anything the online optics have been bad for Anthropic for the last half year. OpenAI doesn't have optics issues, from my point of view they simply have the issue that they are the least trustworthy player at the frontier. The way they pivoted from their original mission is truly breathtaking, especially coming in gloatingly to take the government contract when Anthropic got kicked out for insisting the government does not use their systems for mass surveillance or autonomous weapons systems. You understand what that means, right? OpenAI models are now actively used/developed for mass surveilance and/or autonomous weapons systems.
I know there are plenty here who seem to value their own ability to use these models cheaply above all other considerations. Then OpenAI is a great choice, and much less restrictive than Anthropic. But their problem is not on the optics. It's on the substance.
I really want a good Claude Design competitor in Codex, it's hard to use the others after getting used to it and yet I find anthropic's model to have a much worse understanding of what looks good or not than OpenAI or Google models.
Agreed. GPT 5.5 will come up with more straightforward solutions with far fewer tokens than Claude. Also, the usage limits are much more generous for Codex than Claude Code for the same monthly plan.
Nudged by this thread, I've decided to switch from Claude to Codex for a bit to see what happens. But...I immediately became lost in their marketing vortex of confusion on plans and pricing. Anyone care to tell me which plan I should be using? On the other side I use the $100 Claude Code plan. We actually have a "Business" ChatGPT subscription already, which seems to be $50/mo/seat. OpenAI's web site offers a set of individual subscriptions (for parity with CC presumably) which I suspect weren't available when we signed up for ChatGPT. I think that in turn happened due to some web site feature it didn't allow for free users (uploading PDFs, something like that). Perhaps I should switch from that business account to an individual subscription for Codex?
There is so much less drama involved with the Codex world. You don't realize how oppressive CC is until you've escaped it. Outages, weird restrictions, degradation, accelerated usage, etc etc etc.
Totally. My experience as well. After some time with codex you're like come on Claude can you just stfu! Haha. I now almost always instruct Claude with specific length requirements when I ask questions. Otherwise, it just blathers and blathers in the most annoying of ways. "Oppressive" is spot on in my opinion
I'll agree and expand on "weird restrictions" -- I used to check the claude usage graphs multiple times a day to see where I'm at on my weekly budget. With gpt 5.5 I don't think I'm working differently but haven't felt the need to check anything because I think I've hit my limit... once? on some egregious edge case scenario iirc
Let alone getting banned out right with no reason, zero updates after weeks, and not even being able to download your chat history (despite the feature being available (I assume they vibe coded it and it does not work!). My story below;
I've been using Claude Code, Codex, Gemini (now Antigravity) at the same time for half year now, ever since I dipped my toe into agentic coding. I'd say in general Claude Code and Codex are equally powerful, Gemini is lagging behind.
One thing I appreciate with Codex is, OpenAI nowadays sometimes just gives you quota resets you can bank, so when you use up weekly quota before the week ends, you could just reset the quota, to continue using Codex. I've been much less anxious about Codex quota because of this perk. I just used one reset in the bank yesterday, and still have 3 resets left. Whereas with Claude, when you've used 95% quota 3 days before the week ends, you'd be much more anxious.
On the other hand, Claude Code's /remote-control mechanism is extremely helpful when I am running it in the cloud and wants to monitor it or control it on my phone. Codex currently doesn't support this kind of usage. Codex only allows you to use your phone to connect to a session on your desktop, not in the cloud.
Yes - Anthropic badly needs this same "here's a reset, use it when you want".
It's vastly better this way. Sure, it may impact the bottom line but it's a huge customer satisfaction win.
When Anthropic randomly resets me and I've only used 2%, that's worthless. When OpenAI tells me I have 3 resets available to use whenever I want - it's wonderful.
I have used claude and codex extensively but only from their CLI app (heavily sandboxed using rootless podman, network filtering, etc), so I don't really know what I'm missing with the GUI apps.
One killer feature that Claude has, and AFAIK Codex still lacks, is the ability to start a session in the terminal and then hand it off (actually just remotely control it), from the iOS app.
Last time I tried Codex on iOS it required a ton of set up to link a github project etc. The way claude lets me remote into a session I've already started on my actual machine is much better IMHO.
They’ve addressed that. Codex in the ChatGPT app on iOS is way better than Claude Code now.
You sign in the Codex app on your Mac same on iOS and are able to completely control your sessions - fork, side chats, plugins - everything.
It’s really great i often work through it. And you can connect any number of Codex instances on any number of macs and then manage them all through the iOS app.
I can't tell the difference between Fable and GPT 5.5. I tried Fable while it was in trial $20 mode, used up my whole quota, and it was great, but as soon as I went back to GPT 5.5, everything was the same.
But what I love about Openai is that they still let you hook OTHER harnesses up to a subscription. My Pi setup has been built up for a few months now into exactly what I want and moving over to CC or even Codex is really annoying.
Caveat: I vibe code in tiny little chunks. I see what I want to do, and exactly how I want it done, then prompt that, refine, what was output, then repeat. I bet Fable is better at building a whole app from a 2-sentence prompt; but that's just not important to me at all.
They're different models with different philosophies behind them. This is anecdotal with a user group of 1, but in my experience:
Claude has a stronger personality and is more creative. If you give it vague instructions, it's better at filling in the blanks with reasonable ideas.
GPT-5.5 is better at following instructions. If you know exactly what you want, it will do it without going off the rails. It's also less likely to imply that you're dumb, but I don't really care about that. Some people do.
I’ve found that Claude is very literal. When I talk to 5.5 it gets what i want it to do, when I talk to Opus 4.8 it does what I say literally and doesn’t get the intent behind it.
Claude Code fan here... Codex is very good. Sometimes better. The killer feature is price.
After 6+ months of exclusive Claude Code usage, I was begrudgingly forced to try Codex once they rejiggered their limits such that I kept maxing out my $200/mo plan in just a few days. These days I pay both $200/mo plans, and it's just about enough to get me through a week's work (small game studio - infinite code to write!)
Codex has been good for a long time, more expensive but very focused on efficiency. Working with it feels faster and more to the point than Opus models and I trust it more with long-running jobs. Also regular resets vs being at the whim of Anthropic drama all the time is hella nice.
They've discovered it's a good marketing strategy. Whenever there's an outage, or a new launch, there's often a reset with it, which helps keep people engaged with OAI / Tibo and reduces churn.
They've also introduced banked resets, which are really clever. If you have a $200/month plan and three banked resets, you're not churning because you will overweight giving up those resets (loss aversion theory).
It never really mattered (except when codex was very new). If anything, codex's remote session integration is better, so outside of some "ultracode" orchestration bells/whistles where Claude Code is ahead, I think Codex is a better tool.
> What's the consensus today on codex vs claude code, does it really matter anymore?
Consensus is probably the wrong word for the popular opinions reflected in HN that you might get.
I would recommend that you have 2 of each at all times when it comes to AI so you don't necessarily become overly locked to quirks of one thing. You'll soon realize that things move so fast that you just start internalizing common patterns instead of depending on one specific vendor.
I recommend that you try pi and codex besides claude, to get your own feel for it.
You can however for now use wrappers which are not harnesses such as T3Code though. They were going to cut under the Programmatic API, but have at least temporarily walked it back.
You absolutely can; they are not banning anymore. The bigger problem is that subscription versions of the models are way crappier than when the "same" model is hit via API (Bedrock/Vertex)
You can also make it not count against extra usage.
OpenCode docs show it because Anthropic specifically ambushed them with a PR to remove support so simpletons can't use it easily.
They aren't banning it anymore, they just make it count as "extra usage". e.g. you're paying for every token in addition to your subscription.
Further, the claim that the subscription "version" of the model is worse sounds like bullshit (and the sort of anecdotal nonsense that you see on sites like this). Do you have anything substantiating this?
My experience is that Codex's auto review is extremely costly, with $20 on both sides, I can run CC with auto mode for longer than with Codex's auto review enabled. Also in my own experience Claude's usage is actually bigger than Codex, but I am not sure if that's due to I stick to 5.5 with Codex while keep Sonnet as the default to orchestrate other models in CC.
Not sure about the consensus, but during an entire week I have done every task on my workplace with both Opus 4.8 and GPT 5.5. GPT won hands down. I would even sometimes copy the plans and solutions (using different Git worktrees) from GPT and paste it on Opus and itself would say GPT plans were better. At that point I have migrated. Fable is not enabled in our workspace so I have not tried.
Claude lost my trust around February this year when the plan would say nonsensical things as "delete this method" that was clearly a key method on that part of the codebase.
For personal projects I am using Codex 20$ plan and when that is over I use DeepSeek which is insanely good for the cost.
I use both. Not because I am cool, but because it is cost effective for personal projects with two $20 / month plans. It is also nice to be able to see what the state of the art is like for both.
Personally, I find it very interchangeable. I open codex --yolo or claude with whatever there yolo flag is (have an alias).
More literal, less fluid verbally, harder time understanding nuance, more correct code, fewer bugs. Less pretty UI. I switch back and forth but find I have less 'clean up' work with codex; more upfront communication though to properly specify. High hopes for 5.6!
I personally use opencode so I can swap between models and try different options. I'd say I prefer claude (fable and opus 4.8) so far, but curious to see where gpt 5.6 lands.
For personal stuff, I've been pretty happy with chatgpt's $20 plan. I believe it has considerably higher limits than claude's $20 plan, and it's enough for the personal stuff I play with (hermes, and some small coding stuff). Also allows me to keep up to date on openai models.
I spent the last couple days switching because Anthropic keeps locking stuff behind API pricing. OpenAI lets you do anything with your sub right now. I'm building headless and web interfaces around Pi.dev. I had this previously with Claude Code but they are going to lock away all those features. I think the Claude does a better job at being proactive to solving things, but I'm going to keep tweaking my harness to nudge gpt to do more in it's turn. Not sure!
Personally, I started using openai models to mess with other harnesses. I was pretty oppositional to CC and how they don't let you kinda plug and play freely, or give transparency into -p usage with other harnesses. So i mix and match a bunch of openai and some chinese models im trying out into opencode. I keep hearing codex is great, on the tier of current CC, I've tried it and it just ate my entire 5 hour usage window looping without asking for clarification on something and none of it was usable. that was the only time i tried codex as i could got that same task done with maybe 20% of my window with my existing openai opencode workflow.
I had put a decent amount of effort into setting up that initial codex attempt and it went so poorly that i've been entirely uninterested in trying again. This was maybe a month or so ago, and i know stuff moves fast, but for me, i like the models, dont care for the harness.
If you can afford to test it seriously, running both in parallel, it's worth a test to see which you prefer. If you can't, don't bother. You're not likely missing anything since they are close to personal preference with most people I know who have meaningfully tried both preferring Claude
I use Conductor pretty much exclusively and it makes it incredibly easy to try different models, even within the same workspace - definitely recommend giving it a shot. Whenever I'm forced to use the Claude Code app directly it just seems woefully inadequate compared to Conductor
I sub both codex and claude at 20x. I like opus+fable more than gpt5.5 because it seems gpt tries to finish tasks by leaving any ambiguity unresolved. claude seems better at surfacing open questions.
This is using the same AGENTS.md prompts, which were designed firstly for Claude use, so maybe it's something that could be optimized better if I understood gpt as well?
The answer is it depends. Claude's generally better at frontend and debugging tasks, while Codex is stronger at backend features and exploratory work. They have very different coding styles and thus very different strengths.
Codex app is a much different experience than CC CLI. I would try it out for a couple days with the new model suite and see what you prefer after that.
Last time I tested Codex on a cheap plan, it barely lasted an hour? I think this was for the $20 plan. I was afraid to try the more expensive plan after that. Not sure, I might just outright rip my Claude Code bandaid if the current usage quotas do die off after the 17th or whatever date they said they would "return on".
It's not clear replies to this thread aren't openAI employees or incentivized influencers, but every benchmark has gpt-5.5 underperforming opus 4.8, sometimes by as much as 10%.
If you can afford it and you have something to justify the expense, I would get both. they're interesting to run side by side, you can hand things off from one to the other. Pretty neat. Unfortunately now I just want to have both :(
Not sure there's going to be a consensus, but I can tell you that when i have codex review claude-written code, it finds important gaps and fixes. The reverse is also true. Both are powerful, but even better when used in combination
I left Claude for Codex months ago. I was an early Claude Code adopter but I have found Codex consistently better since about the February time frame. And far more reliable.
It's more diligent and empirical and results focused, and less creative. It sometimes needs a kick to avoid a Zeno's paradox of incremental steps to get to the goal. But it produces more reliable code with fewer race conditions, unhandled negative cases, etc.
It's also better value from a $$ POV, or at least has been. This fluctuates a bit.
You're also free to use your Codex subscription with other harnesses, like opencode, etc. Unlike Anthropic. Plays better with others.
For me the biggest shift was using Deepseek through an American provider with reasonix as the harness, making cache hits at a rate of practically free.
They are both excellent but excel in different areas. Fable is super super proactive and great for doing a LOT of work with a single prompt, also for creative work.
Codex is more details focused, often catches wonky bugs and correctness issues that Fable misses, feels more terse and less "friendly", more like a stern senior engineer versus a friendly talkative engineer (Claude). Codex is also better if you're already an engineer, Claude is better for non-engineers. I.e. Codex works better if you know exactly what you want and know the right way of explaining it.
I figured pi itself would be the best harness because it's barebones and you make it what you want. omp is to pi what doom is to emacs is what lazyvim is to neovim.
The fact that I thought that this was amp misspelled until i someone validate omp and the checked myself indicates it's a subjective assertion at best.
I have one non technical people in my firm using it. One is using it to assist with editing books, basically using it to gather up manuscripts from e-mail / Google Doc etc. submissions, and then switch models between a cheap one and Opus (for actually analysing the manuscript).
The other non-technical person has done really surprising things with it AI, like a long-running GPT 5.5 Pro chat session which is basically her expense tracker - it has an .xlsx file "carried" in the chat, and she just tells ChatGPT (or scans a receipt) whenever she has a new expense, and then prompts it in natural language when she needs a report. I'm looking forward to seeing what she can do with omp.
I've subscribed to ChatGPT/Codex for over a year and tried a Claude sub twice 1 month each, with a gap of several months in between.
I tried them both side by side, mostly for reviewing existing Godot/GDScript code, or sometimes generating Swift Mac apps, including converting ancient artifacts I wrote in Visual Basic on Windows
Besides the useless "This is good" findings while reviewing and the excessive "oops you are right" backtracking, Claude's atrocious UX and borderline "spyware" make me never want to try an Anthropic product again for a long long while.
- the tool calls and results are very legible, I can click them and see the progress
- no one talks about this but the tool call and response notification are handled much more elegantly in Codex. In Claude Code, it is handled in a clunky way using loops which always causes some delay
- you can steer the conversation midway in Codex
- /side is underrated (/btw is the equivalent and is much worse in Claude Code)
- I have to admit subagents are handled better in Claude Code
With the exception of Fable which is going away anyway, Codex is better especially after the last couple Opus releases. It’s also no longer slower than Claude.
You get much more generous usage from the 20x plan.
And you get far better uptime.
If benchmarks and early tester impressions are accurate, you also get access to Fable level capability at greater speed and lower cost (included in subscription).
The fact that they already extended subscription Fable once would suggest it won’t be solely locked behind API next week, but at the same time it really does look like they are doing everything they can to avoid serving it continuously at scale.
Knowing Anthropic, this unfortunately might end up meaning a quietly quantized Fable on subscription.
In my projects, Claude writes and Codex reviews, and I've had a lot of code I've been very happy with out of that, although as of today, Grok _also_ reviews, and finds interesting new stuff.
Pi is so “unbloated” that it’s extra effort to use. You can decide how much work to put into it. I get the trade off. But this is a big jump from CC. I’d recommend some middle ground like opencode.
Even simpler, use Cursor with any frontier model. I see others sweat to add enough context to Claude Code while Cursor has a ton of contextual awareness, uses subagents automatically and is significantly faster with no drop off I have found. I'm not sure why devs are so enamored with living in the CLI, but Cursor has one of those too.
It's worth trying out OpenCode, then oh-my-pi, and also the commercial harnesses like Codex. (I haven't yet bothered to try Antigravity and have no interest in Gemini-cli now that it's not available except on expensive plans.)
pi is also worth tinkering with, particularly if you have an eye towards automating some things.
>> approximately 700,000 A100e GPU hours of black-box automated red teaming
Amusing that they use A100e as the reference point to sound impressive. Different ways you could make that conversion, but based on FP4 FLOPs (yes it's disadvantageous to A100, that's the point), that's something like 200hr on a GB300 NVL72 rack.
Not nothing either, but far less astounding sounding than 700k hrs.
Wait, what do you mean? 700k A100e hours are equal to 200 hours of a GB300 NVL72 rack? One GB300 NVL72, 72-GPU rack has equal processing power to 3500 A100e GPUs?
Funny to see that they did not include Fable 5 in their GeneBench and LifeSciBench comparisons because "it does not answer advanced biology questions and refuses the majority of questions in this eval".
The frontier graph on all these benchmark are extremely in favor of 5.6 Sol over Fable, more than the best model comparisons in previous iterations.
I'd like to know how cherry-picked this is, and what tests it performed less overwhelmingly in, but I suppose that info is not going to be on this post.
If it pans out to be as good as it says, that's great. On the other hand, if this model is not overwhelmingly impressive over Fable, I will lose what remaining trust I had in these announcements.
They do disclose that they scored much lower than Fable on SWEBench Pro, which is a pretty high-quality benchmark. I think it's partially just about what they choose to emphasize...
I totally missed that, because in the charts they showcase for coding, the SWEBench score is not present, they only include it at the end of the post in tables. Hmm.
The SWEBench benchmarks are really gamed at this point and should not be trusted period. The solutions are effectively in the training sets and have been for a while.
The charts are also extremely difficult to parse. They seem auto-generated. Dataset coloring is atrocious.
Regarding your main point, yes, I agree. My impression (as someone who uses both Codex and Claude Code daily) is that OpenAI does a fair amount of benchmaxxing.
I’m interested in knowing how each of GPT 5.6’s variants fare in non-English writing/translation tasks.
GPT 5.5 has a tendency to write English calques and non-idiomatic prose in other languages. Although that can be somewhat tamed with detailed instructions and a corpus of confusing terms, the model’s output often reads like a literal translation rather than native prose. Since I notice these issues most clearly in languages I know well, it makes me reluctant to trust the model’s output in languages in which I’m less proficient.
Ironically, ChatGPT began as a simple text-generation tool, but much of its offerings and benchmarks now focus on coding and agentic workflows, while leaving behind what made it notable in the first place.
"GPT‑5.6 delivers a step change in design judgment. With only high-level direction, GPT‑5.6 creates tasteful, ergonomic, and functional interfaces. Its stronger computer-use capabilities let it inspect and refine the rendered result—not just generate the underlying code or content—so it can catch visual and functional issues and apply finishing touches before handing the work back."
This one is really promising, as it may allow to close major gap with Claude in design/UI skills
Agreed, I’m looking forward to trying it out. I think that the rise of visual design skills that are pretty clearly targeted towards Codex users has lit a bit of a fire under their butts.
GPT usually performs better on DeepSWE while Claude does better on FrontierCode. These two coding benchmarks are pretty much the only ones right now that's still worth taking a look at imo.
DeepSWE seems to strongly, strongly prefer ChatGPT models. There were also major flaws in its methodology pointed out recently, that overlap strongly with the flaws OpenAI pointed out in its SWE Verified report.
I use both ChatGPT and Claude for engineering work on a daily basis, touching performance critical code to application backends to frontend work, and I've found that DeepSWE scores don't reflect my reality when I assess high quality output from the models/harnesses.
Not that Opus always beats GPT 5.5., but that 5.5 is ahead of Opus on a general benchmark smells off to me.
I really wish there was just an easy guide on when to use Sol vs Terra vs Luna, and it just moves further into confusing territory when it comes to naming.
The naming convention is especially difficult to decipher depending on what your native language is. Of course a latin language speaker might be able to easily determine oh yeah each one is slightly bigger than the other but I still think it borderlines too confusing.
That aside all the numbers look amazing, and I'll be happy to probably main this alongside grok-4.5 for a while comparing the two on price and efficiency.
I vastly prefer the direction that OpenAI seems to be going with token efficiency and performance compared to Anthropic who seems to be moving towards a world where you just token-max as much as possible ignoring any and all costs.
Previously it was much more obvious which model to reach for depending on your use case because they had the mini and nano naming conventions.
Getting rid of that seems like a step back. Just a personal nit though.
I've seen buzz about this elsewhere as well but to me effort levels seem more like spend limits disguised with another word. I don't think they should even exist.
The naming convention is bizarre and doesn't really mean anything to normies. Trying to pick between "Sol" and "Terra" is like asking the average person if they want the Max or the Ultra chip.
That isn't what "genuinely asking" looks like, you're criticizing using "questions" as cover. It isn't subtle, nor is it constructive.
I agree with them, Sol, Terra, and Luna are confusing names. They mean the same thing as GPT-5.6-Max, GPT-5.6-Plus, and GPT-5.6-Fast but require base knowledge for an analogy.
It feels like it was adding by the marketing department.
Did you not read the second sentence? Obviously I know what sol is given my first language being Spanish. I'm just speaking in a general sense that it can be confusing for others.
I already know plenty who had no clue what the difference between Terra and Luna would be.
My first instinct was Sol > Luna > Terra, since Sol is the farthest away, then Luna, and Terra is the closest. Size was not my first instinct. Or should Terra be the best model because its closest to people, then Luna because there have been people on it, then Sol be the worst because no human has been there?
Is any of those comparisons about Pro vs non-Pro (Pro is only available in $100+ plans)? I am curious about that but I think Sol, Terra, Luna are different sizes of it without the Pro part, and I want to know how much worse do I have it on the $20 plan compared to if I upgrade.
Oh man, I love capitalism spoiling us here. I was just enjoying my extra Fable credits, now I'll switch to using 5.6 this weekend. I was planning to ration my Anthropic credits, I guess now I do not have to. And I was half wondering if exactly this would happen: right when Fable usage credits were starting to kick in for people, OAI swoops in and takes the puck. As much the AI craze is crazy, this play by play part is pretty fun.
> Instead of requiring developers to script every step or passing every tool response back through the model, Programmatic Tool Calling in the Responses API can filter large amounts of intermediate data, retain only what matters, and adapt its workflow along the way.
Its an extremely capable model. I think the way we need to approach works shifts again. We need to get our harnesses/workflows to let it gather some momentum on the first couple rounds but then we also need to structure it so that it can slingshot and accomplish the long range goal.
I wish model launches were like proper product releases
it's impossible to _try_ it out on release!
it's not on their codex subscription, or the web/mobile chatgpt interfaces, or aws bedrock, etc. I just cant find a working endpoint with the latest model after they announce
The announcement says they're rolling it out over the next 24 hours or so. I think it's reasonable to do a slow-roll-out release for one of the most used products on the internet.
There is an issue on the page that causes the benchmark tables to get cut off. If you highlight and drag right you can see a few more models like Gemini and Claude Opus. It's also interesting that they introduced explicit caching, which is something that only Anthropic had for a long time.
Looks like a great set of models, but there are about 20 different thinking/model levels here in this family and they are very complex to pick the right one for the task
E.g. for GeneBench Pro, it looks like you would always use GPT-5.6 Sol over Terra/Luna, its pareto optimal.
For Agents Last Exam, you would maybe want Luna, then Terra, then Luna, then Sol as you increasingly budget for tasks.
I feel that there may need to be a new auto mode in many of these cases. It selects the best model and thinking given a particular problem.
Feels like it's going to have to go that way eventually, because here we have about 20 different model and thinking levels you could use, and they're not obvious which ones are right for the given use case.
> GPT‑5.6 also introduces more predictable prompt caching, including support for explicit cache breakpoints (opens in a new window) and a 30-minute minimum cache life.
Great to read they are moving away from the 5 minute cache defaults. Hopefully other providers follow soon!
"GPT‑5.6 is available starting today across ChatGPT, Codex, and the OpenAI API. The rollout is starting globally now and will continue gradually toward full availability over the next 24 hours."
My Codex app got upgraded to the new unified ChatGPT app. I don't see Sol available though. Only Terra and Luna. I'm on the Pro plan. Anyone else see it?
"GPT‑5.6 is available starting today across ChatGPT, Codex, and the OpenAI API. The rollout is starting globally now and will continue gradually toward full availability over the next 24 hours."
"On Agents’ Last Exam (opens in a new window), an evaluation of long-running professional workflows across 55 fields, GPT‑5.6 Sol sets a new high of 53.6, eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 points. Even at medium reasoning, it beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost. That efficiency extends to smaller models, which are essential to making intelligence more abundant and affordable: GPT‑5.6 Terra and GPT‑5.6 Luna outperform Fable 5 at around one-sixteenth the cost. "
Some pretty big claims and results! Excited to see how it feels during usage.
I use Fable and 5.5 extensively and I still find both have a place in my toolkit, i.e. Fable IS good but it isn't perfect, and it's still better to play them off against each other. I have Fable and 5.5 write plans and have them adversarially review each other's plans.
Having this amount of competition in the coding model space is good for all of us.
Here's me using a Gemini chat log scraper (from Gdrive) then dumping my prompt+Gemini response into local AI
Never go over the free limits in Gemini Pro.
Gemini is great at research and architecture, and my 30 years experience in programming everything; for fun or work; means together there is little to no code slop.
Add to project repo some git submodules of reference source code; boom, bobs your uncle
Zero reason to sign up for OAI or Claude. With employers realizing the costs are more than employees, local models getting more powerful, and models in chips just a few years out, neither of the one note LLM companies without diversified services and R&D portfolios gonna last
> On the Artificial Analysis Coding Agent Index, GPT‑5.6 Sol with max reasoning sets a new state of the art at 80, 2.8 points above Fable 5, while using less than half the output tokens, taking less than half the time, and costing about one-third less.
> That advantage extends across the family: Terra performs just above Fable 5, while Luna outperforms Opus 4.8; each does so in roughly one-third of the time, with about half as many output tokens, and at approximately one-quarter the estimated cost.
Wow. I don't believe it. Every indication and twitter post told me that Fable is much more intelligent than Sol and here we are told that even Terra outperforms Fable?
Not only that, Sol doesn't even come with run time classifiers. So it is even more suspicious.
What's even stranger is that OpenAI is directly referencing a competitor in this direct way.
SWE-Bench pro is pretty much useless now even though many ppl still look at it. OpenAI published a report yesterday saying so as well. Only look at DeepSWE and FrontierCode right now for coding imo.
SWE-bench series just aren't that great by today's standard, even Anthropic previously stated Claude had memorized solutions for the non Pro version of the benchmark, I suspect the recent increase in the score for the Pro version probably also had similar behaviors.
But anyway, I think it's pretty useless to look at SWE Bench's now when other way better benchmarks exist.
And they'd be right, it's an almost saturated benchmark where even some subpar open source models score very well on. And most models are clustered within a small range so it really doesn't tell you much.
As usual, even though GPT-5.6 is releasing today, the rollout in ChatGPT and Codex will be gradual over many hours so that we can make sure service remains stable for everyone (same as our previous launches). We usually start with Pro/Enterprise accounts and then work our way down to Plus. We know it's slightly annoying to have to wait a random amount of time, but we do it this way to keep service maximally stable.
The timescale is typically hours not minutes, so if you don't see it now, I'd try again later today.
We mention it will be a gradual rollout over the next 24 hours in the Availability section at the bottom of the blog but I admit it's pretty buried.
Is this bug fixed with 5.6? If not, it probably doesn’t matter which version Codex users are getting because the overall result is dramatically worse than stated by Open AI advertising: https://github.com/openai/codex/issues/30364
The marketing team must've done research that said "people are starting to think that you guys are evil-water-stealing-lay-off-loving-bubble-bursting scumbags" and decided to really lean into the small family business and happy font vibes!
Holy shit. They must be feeling very threatened by Fable if they're spending this much energy talking about it in the release notes for their own model.
> GPT‑5.6 is priced per 1M tokens across three model sizes: Sol is $5 input / $30 output; Terra is $2.50 input / $15 output; and Luna is $1 input / $6 output.
Just as expensive as Fable 5. But of course, another slot machine upgrade but the costs will keep going up and the open weight models from china will continue to race everyone else to $0.
Looking forward to the next version of GLM, Qwen, Deepseek and Minimax.
Also watching deepseek closely. Seems like US frontier labs only know how to throw money at things as opposed to actually make smart improvements to the algorithms.
All of them closely collaborate with the government. LLMs are a national security priority and are vetted. Claude AI was used by Palantir's Maven to target the Minab school that led to a triple tap strike killing over 150 schoolchildren.
The developer's guide (https://developers.openai.com/api/docs/guides/latest-model) has some interesting semantic tips for using the model:
> Intent understanding: GPT-5.6 can better infer the user’s underlying goal and intended level of work without you specifying every step. Continue to state important constraints, approval boundaries, and success criteria explicitly.
> Original image detail: GPT-5.6 preserves the original dimensions of images sent with original or auto detail instead of resizing them to a patch budget or pixel-dimension limit.
> Use shorter prompts: In internal evaluations, replacing long, explicit system prompts with minimal prompts improved scores by roughly 10–15%, while reducing total tokens by 41–66% and cost by 33–67%.
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
> Control warmth: GPT-5.6 does not become meaningfully better when prompted to be broadly friendlier or more empathetic.
Control warmth[1]
> GPT-5.6 does not become meaningfully better when prompted to be broadly friendlier or more empathetic. Instead of generic instructions such as “Be friendly and warm,” use concrete guidance: > Be direct and tactful. Acknowledge friction specifically when relevant. Avoid canned reassurance and unnecessary sign-offs.
Soo basically, my new 5.6 custom instructions: Be Jeeves and eliminate all friction from my life through immense processing power. Acknowledge friction specifically when relevant. Avoid canned reassurance and unnecessary sign-offs.
[1] https://developers.openai.com/api/docs/guides/latest-model#c...
> Avoid generic brevity instructions
That part is confusing because it's not like they provide an example of how default GPT-5.6 output compares with GPT-5.5 both with default output and prompted for brevity. Whenever I use such prompts, it's usually because I want the model to give me the gist in a few sentences. I'd be stunned if GPT-5.6 was that concise by default. I would think that could "break" a lot of things for developers who didn't know to make prompt changes after upgrading to 5.6. What if you were expecting GPT to be as wordy as it usually is? Then suddenly your output is not wordy enough?
Smells like OpenAI trying its best to stave off financial armageddon for another few months. Then again, I'm not sure why they chose to waste so much output computation on verbal diarrhea all this time up to now.
It seems like the way brevity instructions have changed is mis-aligned with how most people would expect to use them or are currently using them.
Here's the example they give:
> Instead of asking for the shortest possible answer, replace brevity instructions with prioritization:
> Lead with the conclusion. Include the evidence needed to support it, any material caveat, and the next action. Omit secondary detail and repetition.
> Keep all required facts, decisions, caveats, and next steps. Trim introductions, repetition, generic reassurance, and optional background first.
Generally speaking, when I ask for a short answer, I want a short answer because I'm not really willing to read through a bunch of bullshit to get to a summary. Putting the onus back on me to assume what the model will return and write a longer prompt detailing exactly what information I want completely misses the point of why I'm asking for a short answer in the first place.
Replace 2 word instruction ('be concise') with a 38 word instruction.
Human can no longer be concise when asking for a few sentences instead of 20 paragraphs of BS they don't want to read when all they want is a summary to verify the general direction of the prompt-work before digging into the details.
such progress!
I don't know how intentional it is / was, but LLMs in general just love to hear themselves talk!
They do, and I want to encourage them to do so because they think through talking. What I don’t want to do is spend time reading all that.
We will probably just get reader-side affordances for this like auto-folded justification and introduction sections and so on.
Doubtless some chat interface will add this the way they’ve added reasoning folding.
Thinking models think through talking, don't reveal that talking, then answer by again thinking through talking. It's kinda funny in a way.
Is it just a coincidence that the companies creating them charge by the token?
The aligned incentive appears to be realigning in favor of the corporation.
Pray they do not realign them further.
There are times I require single word answers. I will use whatever model responds as I desire and at this point those models are just a few.
do we have similar guidance or page from anthropic for claude?
> Use shorter prompts: In internal evaluations, replacing long, explicit system prompts with minimal prompts improved scores by roughly 10–15%, while reducing total tokens by 41–66% and cost by 33–67%.
When has this ever not been the case? I don't think this is a GPT 5.6 specialty!
Information density of the prompt is the most important factor in my experience.
And interestingly, LLMs seem particularly bad at writing prompts for other LLMs for this reason (you can guide them to be more dense, just speaking by default).
Conciseness is usually a byproduct of information density though.
> Use shorter prompts: In internal evaluations, replacing long, explicit system prompts with minimal prompts improved scores by roughly 10–15%, while reducing total tokens by 41–66% and cost by 33–67%.
A shorter prompt results in half as much tokens spend? I find this very hard to believe.
If it's anywhere close to the same universe as smaller models in its behavior, a lot of time in "thinking" mode is spent on reiterating on any constraints given in a prompt. So the more constraints you give it, the more tokens it will spend going "Hold on, the prompt said I have to dot my i's and cross my t's. Let me go through my work to check that all the i's are dotted."
Maybe Codex has the same problem I sometimes have focusing while reading and has to reread the same sentence over and over again.
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
What about my favorite, "no yapping"?
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
RIP Caveman skill. Six month good. Now skill dead.
A Yoda skill, is there?
GPT-5.6 Sol sets a new SOTA on ARC-AGI-3: 7.8%
Sol is the first verified frontier model to ever beat an ARC-AGI-3 game
https://arcprize.org/results/openai-gpt-5-6
Seeing the dramatic differences in scores just going from high to xhigh is just another demonstration of the bitter lesson: Just keep scaling search and learning. We are probably going to need a lot more GPUs.
While I think this is true, remember as we get more efficient we just decide to scale even bigger. So more GPUs, and more efficient.
I agree with the sibling comment, effiency is probably the more important component at this point. We are hitting not just a practical engineering roadblock for scaling with current technology, I think we have definitely hit a financial and logistical roadblock for up scaling with the number of GPUs (on an immediate basis)
I mean, theoretically you can solve every finitary problem with a brute force solution...
Richard Sutton specifically states that the search has to be smart. We know that the brain uses recurrent connections and is shallow. I think a lot more money has to go into architecture. Feed Forward transformers can only scale so far
> We are probably going to need a lot more GPUs.
Or a breakthrough in algorithms etc.
The human brain, heck all bio brains, are proof that you don't need a lot of power or size for intelligence.
Very interesting. My prediction is that Mythos would outperform Sol.
Also what does this tell about Yann LeCuns whole world model theory? Bro has been going on and on about it. He has made multiple wrong predictions on the trajectory of LLMs.
At some point his claim should be fully falsified no?
Mythos probably wouldn't, otherwise they'd have included it in their release. Next version of Mythos probably will though.
And yeah.. Reality has not been kind to LeCun.
Are you joking? They spend billions of dollars training LLMs to get a 7.8% on arc agi 3 whereas DINO models are near sota in image classification, provide meaningful embeddings to the point where image segmentation is just PCA. The spend on DINO cannot be more than five million (correct me if I'm wrong)
JEPA is just getting started
His main strong anti-LLM predictions have been consistently either wrong or misleading.
There's many ways to skin a cat so you can probably do something with a JEPA approach as well, but I doubt he actually catches up to having agents on the level of where Anthropic/OpenAI will be at any point.
DINO is a transformer model?
ASI is going to be here by the time Lecun gets started.
Ok long time Claude Code user here; lately I've started to realize there's other great models out there I should be trying, but I'm hesitant to leave Claude Code behind for something new.
What's the consensus today on codex vs claude code, does it really matter anymore?
Codex has arguably been better than Claude Code for months now, but it's flown under the radar because it just didn't capture the same viral marketing effect and OpenAI in general has had more optics / PR issues than Anthropic amongst the online developer crowd. I use the word "better" not in the sense that the underlying GPT models are fundamentally smarter or more intelligent, but rather that as a product Codex is just simpler, cheaper, and abundantly reliable and low-drama.
I’d argue the opposite. I’ve switched back and forth from one to the other and Opus/Fable has been constantly better than any GPT in my daily work. It’s a bit slower but it does the things right, with as little code as possible, some comments where needed. Codex is faster but you always have to correct it because it got something wrong; it writes tons of code ("let me add a small helper") with obvious comments.
Purely anecdotally the one persistent issue I have with LLMs writing code is that they are absolutely paranoid and add a load of indirection and defensive crap and even if you prompt to avoid that it will often require manual steering to remove the cruft.
I really love the Opus/Fable models but I'm honestly sick to death of the buggy product. The CLI always has some weird issue. Right now it doesn't even output messages before tool calls, it just swallows them and they disappear.
I don't like OpenAI as a company, but they appear to have QA, and that is probably enough to get me to switch.
There was an issue on Claude Code the other day where it would only wait 60 seconds when it had asked a set of questions, then if it didn't get a response from the user it would just continue however it thought was best. Completely unusable. It took them nearly 48 hours to merge a fix.
I'm not sure how meaningful this is. Fable only just recently become more broadly available, and GPT-5.6 is launching broadly today.
Sounds like you are talking past each other. GP is saying the harness of codex is higher quality, which I can believe, even if the models are not as good as Opus/Fable.
That's a strange statement... It's been true for a while now that OpenAI has had much more generous limits than Anthropic on their subscription plans. And with the Fable ban/guardrails disaster, there has been a lot of frustration from people in these comment sections. And Anthropic fucked up Claude Code pretty badly for a couple of weeks during the 4.6/4.7/4.8 transition, which again was widely publicized. And they got a lot of flack over not allowing other harnesses anymore. And ChatGPT got some pretty viral wins on model intelligence when they cracked the high profile Erdos problem.
If anything the online optics have been bad for Anthropic for the last half year. OpenAI doesn't have optics issues, from my point of view they simply have the issue that they are the least trustworthy player at the frontier. The way they pivoted from their original mission is truly breathtaking, especially coming in gloatingly to take the government contract when Anthropic got kicked out for insisting the government does not use their systems for mass surveillance or autonomous weapons systems. You understand what that means, right? OpenAI models are now actively used/developed for mass surveilance and/or autonomous weapons systems.
I know there are plenty here who seem to value their own ability to use these models cheaply above all other considerations. Then OpenAI is a great choice, and much less restrictive than Anthropic. But their problem is not on the optics. It's on the substance.
I really want a good Claude Design competitor in Codex, it's hard to use the others after getting used to it and yet I find anthropic's model to have a much worse understanding of what looks good or not than OpenAI or Google models.
Agreed. GPT 5.5 will come up with more straightforward solutions with far fewer tokens than Claude. Also, the usage limits are much more generous for Codex than Claude Code for the same monthly plan.
Last time I used Codex it would make loads of assumptions, often quite big ones, without asking.
Did they fix that, as that for me was what actually made codex worse.
Nudged by this thread, I've decided to switch from Claude to Codex for a bit to see what happens. But...I immediately became lost in their marketing vortex of confusion on plans and pricing. Anyone care to tell me which plan I should be using? On the other side I use the $100 Claude Code plan. We actually have a "Business" ChatGPT subscription already, which seems to be $50/mo/seat. OpenAI's web site offers a set of individual subscriptions (for parity with CC presumably) which I suspect weren't available when we signed up for ChatGPT. I think that in turn happened due to some web site feature it didn't allow for free users (uploading PDFs, something like that). Perhaps I should switch from that business account to an individual subscription for Codex?
There is so much less drama involved with the Codex world. You don't realize how oppressive CC is until you've escaped it. Outages, weird restrictions, degradation, accelerated usage, etc etc etc.
Totally. My experience as well. After some time with codex you're like come on Claude can you just stfu! Haha. I now almost always instruct Claude with specific length requirements when I ask questions. Otherwise, it just blathers and blathers in the most annoying of ways. "Oppressive" is spot on in my opinion
I'll agree and expand on "weird restrictions" -- I used to check the claude usage graphs multiple times a day to see where I'm at on my weekly budget. With gpt 5.5 I don't think I'm working differently but haven't felt the need to check anything because I think I've hit my limit... once? on some egregious edge case scenario iirc
Let alone getting banned out right with no reason, zero updates after weeks, and not even being able to download your chat history (despite the feature being available (I assume they vibe coded it and it does not work!). My story below;
https://news.ycombinator.com/item?id=48597861
Even less drama with open models like GLM.
I've been using Claude Code, Codex, Gemini (now Antigravity) at the same time for half year now, ever since I dipped my toe into agentic coding. I'd say in general Claude Code and Codex are equally powerful, Gemini is lagging behind.
One thing I appreciate with Codex is, OpenAI nowadays sometimes just gives you quota resets you can bank, so when you use up weekly quota before the week ends, you could just reset the quota, to continue using Codex. I've been much less anxious about Codex quota because of this perk. I just used one reset in the bank yesterday, and still have 3 resets left. Whereas with Claude, when you've used 95% quota 3 days before the week ends, you'd be much more anxious.
On the other hand, Claude Code's /remote-control mechanism is extremely helpful when I am running it in the cloud and wants to monitor it or control it on my phone. Codex currently doesn't support this kind of usage. Codex only allows you to use your phone to connect to a session on your desktop, not in the cloud.
Yes - Anthropic badly needs this same "here's a reset, use it when you want".
It's vastly better this way. Sure, it may impact the bottom line but it's a huge customer satisfaction win.
When Anthropic randomly resets me and I've only used 2%, that's worthless. When OpenAI tells me I have 3 resets available to use whenever I want - it's wonderful.
Codex is supported well on iPhone/iPad, it’s inside the ChatGPT app.
It’s amazing how much work you can get done on your phone now, especially if you already have a design mapped out in your head.
I have used claude and codex extensively but only from their CLI app (heavily sandboxed using rootless podman, network filtering, etc), so I don't really know what I'm missing with the GUI apps.
One killer feature that Claude has, and AFAIK Codex still lacks, is the ability to start a session in the terminal and then hand it off (actually just remotely control it), from the iOS app.
Last time I tried Codex on iOS it required a ton of set up to link a github project etc. The way claude lets me remote into a session I've already started on my actual machine is much better IMHO.
They’ve addressed that. Codex in the ChatGPT app on iOS is way better than Claude Code now.
You sign in the Codex app on your Mac same on iOS and are able to completely control your sessions - fork, side chats, plugins - everything.
It’s really great i often work through it. And you can connect any number of Codex instances on any number of macs and then manage them all through the iOS app.
I've been using codex app server. Works great.
https://learn.chatgpt.com/docs/app-server
I can't tell the difference between Fable and GPT 5.5. I tried Fable while it was in trial $20 mode, used up my whole quota, and it was great, but as soon as I went back to GPT 5.5, everything was the same.
But what I love about Openai is that they still let you hook OTHER harnesses up to a subscription. My Pi setup has been built up for a few months now into exactly what I want and moving over to CC or even Codex is really annoying.
Caveat: I vibe code in tiny little chunks. I see what I want to do, and exactly how I want it done, then prompt that, refine, what was output, then repeat. I bet Fable is better at building a whole app from a 2-sentence prompt; but that's just not important to me at all.
Same here - gave 5.5 a web design to implement and it sucked. Gave the same to Fable and it still sucked.
> does it really matter anymore?
They're different models with different philosophies behind them. This is anecdotal with a user group of 1, but in my experience:
Claude has a stronger personality and is more creative. If you give it vague instructions, it's better at filling in the blanks with reasonable ideas.
GPT-5.5 is better at following instructions. If you know exactly what you want, it will do it without going off the rails. It's also less likely to imply that you're dumb, but I don't really care about that. Some people do.
I’ve found that Claude is very literal. When I talk to 5.5 it gets what i want it to do, when I talk to Opus 4.8 it does what I say literally and doesn’t get the intent behind it.
Claude Code fan here... Codex is very good. Sometimes better. The killer feature is price.
After 6+ months of exclusive Claude Code usage, I was begrudgingly forced to try Codex once they rejiggered their limits such that I kept maxing out my $200/mo plan in just a few days. These days I pay both $200/mo plans, and it's just about enough to get me through a week's work (small game studio - infinite code to write!)
Codex has been good for a long time, more expensive but very focused on efficiency. Working with it feels faster and more to the point than Opus models and I trust it more with long-running jobs. Also regular resets vs being at the whim of Anthropic drama all the time is hella nice.
Codex is cheaper on average no? I think the models are expensive but the token efficiency of the harness itself solves the problem.
Yes that's what I meant, the per token cost is higher but as you say the efficiency levels it out / works slightly in Codex's favour.
Anyone know what the deal is with the resets?
They've discovered it's a good marketing strategy. Whenever there's an outage, or a new launch, there's often a reset with it, which helps keep people engaged with OAI / Tibo and reduces churn.
They've also introduced banked resets, which are really clever. If you have a $200/month plan and three banked resets, you're not churning because you will overweight giving up those resets (loss aversion theory).
I ran out of resets :( hehe I had 3 and used them all
It never really mattered (except when codex was very new). If anything, codex's remote session integration is better, so outside of some "ultracode" orchestration bells/whistles where Claude Code is ahead, I think Codex is a better tool.
Agree, I think there was just a blind study that showed no one could tell the difference even though the users were avid they could
just try it you will back to codex because gpt is trash, I ask for refund under 7 hours
> What's the consensus today on codex vs claude code, does it really matter anymore?
Consensus is probably the wrong word for the popular opinions reflected in HN that you might get.
I would recommend that you have 2 of each at all times when it comes to AI so you don't necessarily become overly locked to quirks of one thing. You'll soon realize that things move so fast that you just start internalizing common patterns instead of depending on one specific vendor.
I recommend that you try pi and codex besides claude, to get your own feel for it.
Use a harness that doesn't lock you into a moat, like OpenCode.
Codex is open source and lets you use any model https://learn.chatgpt.com/docs/config-file/config-advanced#o...
FWIW Claude Code works with OpenRouter so you can use any model.
CC system prompt is bloated, use Pi to test Codex instead
Codex CLI is open source too. I don't think there is a difference.
Can't use a claude code subscription in another harness though
Don't use providers that don't allow it.
You can however for now use wrappers which are not harnesses such as T3Code though. They were going to cut under the Programmatic API, but have at least temporarily walked it back.
You absolutely can; they are not banning anymore. The bigger problem is that subscription versions of the models are way crappier than when the "same" model is hit via API (Bedrock/Vertex)
You can also make it not count against extra usage.
OpenCode docs show it because Anthropic specifically ambushed them with a PR to remove support so simpletons can't use it easily.
The opencode docs[0] still say otherwise, do you have a source?
[0] https://opencode.ai/docs/providers/#anthropic
I am curious about the claim that the subscription models are different. Has anyone benchmarked this?
Do you have a source for that?
They aren't banning it anymore, they just make it count as "extra usage". e.g. you're paying for every token in addition to your subscription.
Further, the claim that the subscription "version" of the model is worse sounds like bullshit (and the sort of anecdotal nonsense that you see on sites like this). Do you have anything substantiating this?
My experience is that Codex's auto review is extremely costly, with $20 on both sides, I can run CC with auto mode for longer than with Codex's auto review enabled. Also in my own experience Claude's usage is actually bigger than Codex, but I am not sure if that's due to I stick to 5.5 with Codex while keep Sonnet as the default to orchestrate other models in CC.
Not sure about the consensus, but during an entire week I have done every task on my workplace with both Opus 4.8 and GPT 5.5. GPT won hands down. I would even sometimes copy the plans and solutions (using different Git worktrees) from GPT and paste it on Opus and itself would say GPT plans were better. At that point I have migrated. Fable is not enabled in our workspace so I have not tried.
Claude lost my trust around February this year when the plan would say nonsensical things as "delete this method" that was clearly a key method on that part of the codebase.
For personal projects I am using Codex 20$ plan and when that is over I use DeepSeek which is insanely good for the cost.
I use both. Not because I am cool, but because it is cost effective for personal projects with two $20 / month plans. It is also nice to be able to see what the state of the art is like for both.
Personally, I find it very interchangeable. I open codex --yolo or claude with whatever there yolo flag is (have an alias).
More literal, less fluid verbally, harder time understanding nuance, more correct code, fewer bugs. Less pretty UI. I switch back and forth but find I have less 'clean up' work with codex; more upfront communication though to properly specify. High hopes for 5.6!
They blocked Claude from being used in a different harness as well squeezed the usage like crazy. Switched to Codex and haven't cared since.
Between the two the biggest difference by far is ... getting your harness / AGENTS.md / skills / tools set up right.
I personally use opencode so I can swap between models and try different options. I'd say I prefer claude (fable and opus 4.8) so far, but curious to see where gpt 5.6 lands.
For personal stuff, I've been pretty happy with chatgpt's $20 plan. I believe it has considerably higher limits than claude's $20 plan, and it's enough for the personal stuff I play with (hermes, and some small coding stuff). Also allows me to keep up to date on openai models.
The $20 GPT plan with GPT 5.5 lasted me, somehow, exactly one smallish fixup feature
I spent the last couple days switching because Anthropic keeps locking stuff behind API pricing. OpenAI lets you do anything with your sub right now. I'm building headless and web interfaces around Pi.dev. I had this previously with Claude Code but they are going to lock away all those features. I think the Claude does a better job at being proactive to solving things, but I'm going to keep tweaking my harness to nudge gpt to do more in it's turn. Not sure!
Personally, I started using openai models to mess with other harnesses. I was pretty oppositional to CC and how they don't let you kinda plug and play freely, or give transparency into -p usage with other harnesses. So i mix and match a bunch of openai and some chinese models im trying out into opencode. I keep hearing codex is great, on the tier of current CC, I've tried it and it just ate my entire 5 hour usage window looping without asking for clarification on something and none of it was usable. that was the only time i tried codex as i could got that same task done with maybe 20% of my window with my existing openai opencode workflow.
I had put a decent amount of effort into setting up that initial codex attempt and it went so poorly that i've been entirely uninterested in trying again. This was maybe a month or so ago, and i know stuff moves fast, but for me, i like the models, dont care for the harness.
If you can afford to test it seriously, running both in parallel, it's worth a test to see which you prefer. If you can't, don't bother. You're not likely missing anything since they are close to personal preference with most people I know who have meaningfully tried both preferring Claude
In my experience, for coding Codex is definitely far ahead of Claude Code, even when using Fable 5 as a model.
you have a very strange experience
I use Conductor pretty much exclusively and it makes it incredibly easy to try different models, even within the same workspace - definitely recommend giving it a shot. Whenever I'm forced to use the Claude Code app directly it just seems woefully inadequate compared to Conductor
I sub both codex and claude at 20x. I like opus+fable more than gpt5.5 because it seems gpt tries to finish tasks by leaving any ambiguity unresolved. claude seems better at surfacing open questions.
This is using the same AGENTS.md prompts, which were designed firstly for Claude use, so maybe it's something that could be optimized better if I understood gpt as well?
The answer is it depends. Claude's generally better at frontend and debugging tasks, while Codex is stronger at backend features and exploratory work. They have very different coding styles and thus very different strengths.
Any actual data backing this up? Or is this just your personal experience?
Codex app is a much different experience than CC CLI. I would try it out for a couple days with the new model suite and see what you prefer after that.
Last time I tested Codex on a cheap plan, it barely lasted an hour? I think this was for the $20 plan. I was afraid to try the more expensive plan after that. Not sure, I might just outright rip my Claude Code bandaid if the current usage quotas do die off after the 17th or whatever date they said they would "return on".
I have found Claude Code to be so much better than other common harnesses that it's kept me solely in the Anthropic ecosystem.
It's not clear replies to this thread aren't openAI employees or incentivized influencers, but every benchmark has gpt-5.5 underperforming opus 4.8, sometimes by as much as 10%.
Can they all be wrong/paid-off?
You wouldn't be leaving Claude Code, just trying something new. If you don't like it just resume using Claude.
If you can afford it and you have something to justify the expense, I would get both. they're interesting to run side by side, you can hand things off from one to the other. Pretty neat. Unfortunately now I just want to have both :(
Not sure there's going to be a consensus, but I can tell you that when i have codex review claude-written code, it finds important gaps and fixes. The reverse is also true. Both are powerful, but even better when used in combination
IMO LMArena is the best benchmark that avoids benchmaxxing
https://arena.ai/leaderboard/agent
5.6 isn’t on there yet but Fable leads by a significant margin atm
The results here match up to my real-world experience using these models every day at work and switching between them regularly.
There was just a study showing that when presented blindly no one could tell the difference yet users were avid they could
There _is_ a difference in the way Claude and GPT write. Last Friday I felt Opus was becoming dumb because it was writing like GPT.
I left Claude for Codex months ago. I was an early Claude Code adopter but I have found Codex consistently better since about the February time frame. And far more reliable.
It's more diligent and empirical and results focused, and less creative. It sometimes needs a kick to avoid a Zeno's paradox of incremental steps to get to the goal. But it produces more reliable code with fewer race conditions, unhandled negative cases, etc.
It's also better value from a $$ POV, or at least has been. This fluctuates a bit.
You're also free to use your Codex subscription with other harnesses, like opencode, etc. Unlike Anthropic. Plays better with others.
For me the biggest shift was using Deepseek through an American provider with reasonix as the harness, making cache hits at a rate of practically free.
I wish they open source their desktop app and built-in skills one day. That would be a final blow for me.
I use both especially for checking each others work. Pretty happy with results
They are both excellent but excel in different areas. Fable is super super proactive and great for doing a LOT of work with a single prompt, also for creative work.
Codex is more details focused, often catches wonky bugs and correctness issues that Fable misses, feels more terse and less "friendly", more like a stern senior engineer versus a friendly talkative engineer (Claude). Codex is also better if you're already an engineer, Claude is better for non-engineers. I.e. Codex works better if you know exactly what you want and know the right way of explaining it.
Consensus itself does NOT matter, omp is objectively the best harness for power users yet it has 0 hn posts about it, zero.
You're fully free to use and try anything and without caring about what others think is right
I figured pi itself would be the best harness because it's barebones and you make it what you want. omp is to pi what doom is to emacs is what lazyvim is to neovim.
The fact that I thought that this was amp misspelled until i someone validate omp and the checked myself indicates it's a subjective assertion at best.
omp is really good.
I have one non technical people in my firm using it. One is using it to assist with editing books, basically using it to gather up manuscripts from e-mail / Google Doc etc. submissions, and then switch models between a cheap one and Opus (for actually analysing the manuscript).
The other non-technical person has done really surprising things with it AI, like a long-running GPT 5.5 Pro chat session which is basically her expense tracker - it has an .xlsx file "carried" in the chat, and she just tells ChatGPT (or scans a receipt) whenever she has a new expense, and then prompts it in natural language when she needs a report. I'm looking forward to seeing what she can do with omp.
Absolutely. It's the only harness that is actually RSI and not run by idiots.
How can this be "objective"? Surely its subjective.
I've tried a fuck load of harnesses but keep coming back to Codex as my harness.
> omp is objectively the best harness for power users
Care to detail this?
"objectively the best"?
Probably means subjectively according to his own opinion...
Try OpenCode and you can point it at either model
I've subscribed to ChatGPT/Codex for over a year and tried a Claude sub twice 1 month each, with a gap of several months in between.
I tried them both side by side, mostly for reviewing existing Godot/GDScript code, or sometimes generating Swift Mac apps, including converting ancient artifacts I wrote in Visual Basic on Windows
Codex was consistently better than Claude: https://i.imgur.com/jYawPDY.png
Besides the useless "This is good" findings while reviewing and the excessive "oops you are right" backtracking, Claude's atrocious UX and borderline "spyware" make me never want to try an Anthropic product again for a long long while.
Codex UI is way way way better than Claude Code
- codex UI is much more responsive
- i get feedback about the progress easily
- the tool calls and results are very legible, I can click them and see the progress
- no one talks about this but the tool call and response notification are handled much more elegantly in Codex. In Claude Code, it is handled in a clunky way using loops which always causes some delay
- you can steer the conversation midway in Codex
- /side is underrated (/btw is the equivalent and is much worse in Claude Code)
- I have to admit subagents are handled better in Claude Code
With the exception of Fable which is going away anyway, Codex is better especially after the last couple Opus releases. It’s also no longer slower than Claude.
You get much more generous usage from the 20x plan.
And you get far better uptime.
If benchmarks and early tester impressions are accurate, you also get access to Fable level capability at greater speed and lower cost (included in subscription).
> Fable which is going away anyway
$2 says nah. You can't take Fable away in a week where GPT-5.6 and Grok 4.5 launch, if you want to hold on to customers.
The fact that they already extended subscription Fable once would suggest it won’t be solely locked behind API next week, but at the same time it really does look like they are doing everything they can to avoid serving it continuously at scale.
Knowing Anthropic, this unfortunately might end up meaning a quietly quantized Fable on subscription.
I use Claude for planning, writing CRs, and code review.
Codex writes all of the code, no exceptions.
Works great, especially when you ask Claude to break up large CRs into roughly 10 minutes of Codex work each.
> I'm hesitant to leave Claude Code behind for something new.
Codex and Claude Code are not mutually exclusive, you can use both.
In my projects, Claude writes and Codex reviews, and I've had a lot of code I've been very happy with out of that, although as of today, Grok _also_ reviews, and finds interesting new stuff.
I use both constantly for different things. You don't need to be a one-model Andy
Claude Code is a massively bloated agent harness.
Try Pi: https://pi.dev/
Pi is so “unbloated” that it’s extra effort to use. You can decide how much work to put into it. I get the trade off. But this is a big jump from CC. I’d recommend some middle ground like opencode.
Even simpler, use Cursor with any frontier model. I see others sweat to add enough context to Claude Code while Cursor has a ton of contextual awareness, uses subagents automatically and is significantly faster with no drop off I have found. I'm not sure why devs are so enamored with living in the CLI, but Cursor has one of those too.
It's worth trying out OpenCode, then oh-my-pi, and also the commercial harnesses like Codex. (I haven't yet bothered to try Antigravity and have no interest in Gemini-cli now that it's not available except on expensive plans.)
pi is also worth tinkering with, particularly if you have an eye towards automating some things.
Literally every top model is identical and anyone saying otherwise is engaging in astrology.
The outputs, ui, and overall behavior (tokenization) are not identical.
anybody saying they're identical clearly doesn't use both...
Honestly, I would even push it further. People who would claim that don't use either one.
>> approximately 700,000 A100e GPU hours of black-box automated red teaming
Amusing that they use A100e as the reference point to sound impressive. Different ways you could make that conversion, but based on FP4 FLOPs (yes it's disadvantageous to A100, that's the point), that's something like 200hr on a GB300 NVL72 rack.
Not nothing either, but far less astounding sounding than 700k hrs.
Wait, what do you mean? 700k A100e hours are equal to 200 hours of a GB300 NVL72 rack? One GB300 NVL72, 72-GPU rack has equal processing power to 3500 A100e GPUs?
Funny to see that they did not include Fable 5 in their GeneBench and LifeSciBench comparisons because "it does not answer advanced biology questions and refuses the majority of questions in this eval".
Winner by default!
This is a major reason why I and a number of biologists I've talked to have canceled their anthropic accounts recently. Not working is not working.
Where’s the lie?
The frontier graph on all these benchmark are extremely in favor of 5.6 Sol over Fable, more than the best model comparisons in previous iterations.
I'd like to know how cherry-picked this is, and what tests it performed less overwhelmingly in, but I suppose that info is not going to be on this post.
If it pans out to be as good as it says, that's great. On the other hand, if this model is not overwhelmingly impressive over Fable, I will lose what remaining trust I had in these announcements.
They do disclose that they scored much lower than Fable on SWEBench Pro, which is a pretty high-quality benchmark. I think it's partially just about what they choose to emphasize...
It's worth noting that OpenAI recently came out saying, "We don't think SWEBench Pro is worth reporting any more" - https://openai.com/index/separating-signal-from-noise-coding...
lol
SWEBench Pro should be ignored until they fix it or disprove the broken task accusations.
I totally missed that, because in the charts they showcase for coding, the SWEBench score is not present, they only include it at the end of the post in tables. Hmm.
Great catch.
The SWEBench benchmarks are really gamed at this point and should not be trusted period. The solutions are effectively in the training sets and have been for a while.
Didn't they also just post about how SWEBench is broken?
> SWEBench Pro, which is a pretty high-quality benchmark
No, doesn't seem like it
https://openai.com/index/separating-signal-from-noise-coding...
The proof is in the pudding and these benchmark stats will only work for so long before people lose interest.
The charts are also extremely difficult to parse. They seem auto-generated. Dataset coloring is atrocious.
Regarding your main point, yes, I agree. My impression (as someone who uses both Codex and Claude Code daily) is that OpenAI does a fair amount of benchmaxxing.
I’m interested in knowing how each of GPT 5.6’s variants fare in non-English writing/translation tasks.
GPT 5.5 has a tendency to write English calques and non-idiomatic prose in other languages. Although that can be somewhat tamed with detailed instructions and a corpus of confusing terms, the model’s output often reads like a literal translation rather than native prose. Since I notice these issues most clearly in languages I know well, it makes me reluctant to trust the model’s output in languages in which I’m less proficient.
Ironically, ChatGPT began as a simple text-generation tool, but much of its offerings and benchmarks now focus on coding and agentic workflows, while leaving behind what made it notable in the first place.
"GPT‑5.6 delivers a step change in design judgment. With only high-level direction, GPT‑5.6 creates tasteful, ergonomic, and functional interfaces. Its stronger computer-use capabilities let it inspect and refine the rendered result—not just generate the underlying code or content—so it can catch visual and functional issues and apply finishing touches before handing the work back."
This one is really promising, as it may allow to close major gap with Claude in design/UI skills
Agreed, I’m looking forward to trying it out. I think that the rise of visual design skills that are pretty clearly targeted towards Codex users has lit a bit of a fire under their butts.
"We've extended usage of Claude Fable" message incoming any day now.
They reset all usage half an hour ago. It's back to 0% per week and session. No specifically Fable related.
Hahaha seeing this play out in real time is absolutely incredible.
Wow the video is much better.. the PR spend clearly went up a lot. Mainly just showing "real people" doing "real stuff".
5.6 Terra (mid tier model) as good as Fable on DeepSWE while cheaper than Opus API pricing. Seems like a homerun.
GPT usually performs better on DeepSWE while Claude does better on FrontierCode. These two coding benchmarks are pretty much the only ones right now that's still worth taking a look at imo.
DeepSWE seems to strongly, strongly prefer ChatGPT models. There were also major flaws in its methodology pointed out recently, that overlap strongly with the flaws OpenAI pointed out in its SWE Verified report.
I use both ChatGPT and Claude for engineering work on a daily basis, touching performance critical code to application backends to frontend work, and I've found that DeepSWE scores don't reflect my reality when I assess high quality output from the models/harnesses.
Not that Opus always beats GPT 5.5., but that 5.5 is ahead of Opus on a general benchmark smells off to me.
I haven't tried an OpenAI model for a long time, but with Fable going to API pricing soon this might be enough to get me to try codex.
I really wish there was just an easy guide on when to use Sol vs Terra vs Luna, and it just moves further into confusing territory when it comes to naming.
The naming convention is especially difficult to decipher depending on what your native language is. Of course a latin language speaker might be able to easily determine oh yeah each one is slightly bigger than the other but I still think it borderlines too confusing.
That aside all the numbers look amazing, and I'll be happy to probably main this alongside grok-4.5 for a while comparing the two on price and efficiency.
I vastly prefer the direction that OpenAI seems to be going with token efficiency and performance compared to Anthropic who seems to be moving towards a world where you just token-max as much as possible ignoring any and all costs.
Why would you need a guide for that now? We long had to pick different models (and thinking levels) by task and feel.
Previously it was much more obvious which model to reach for depending on your use case because they had the mini and nano naming conventions.
Getting rid of that seems like a step back. Just a personal nit though.
I've seen buzz about this elsewhere as well but to me effort levels seem more like spend limits disguised with another word. I don't think they should even exist.
The naming convention is bizarre and doesn't really mean anything to normies. Trying to pick between "Sol" and "Terra" is like asking the average person if they want the Max or the Ultra chip.
What about Haiku, Sonnet, and Opus?
The sun is bigger than earth which is bigger than the moon, it's pretty simple really
Which is cheaper to use? The size euphemism is a really roundabout description vs "Nano" and "Pro" for the layperson.
it's simple: unless trivial TOIL, always use the highest at ultra max settings.
Okay Richie Rich
You don’t know what sol means? You don’t understand the difference in sizes between Terra and sol? I’m genuinely asking.
That isn't what "genuinely asking" looks like, you're criticizing using "questions" as cover. It isn't subtle, nor is it constructive.
I agree with them, Sol, Terra, and Luna are confusing names. They mean the same thing as GPT-5.6-Max, GPT-5.6-Plus, and GPT-5.6-Fast but require base knowledge for an analogy.
It feels like it was adding by the marketing department.
Did you not read the second sentence? Obviously I know what sol is given my first language being Spanish. I'm just speaking in a general sense that it can be confusing for others.
I already know plenty who had no clue what the difference between Terra and Luna would be.
My first instinct was Sol > Luna > Terra, since Sol is the farthest away, then Luna, and Terra is the closest. Size was not my first instinct. Or should Terra be the best model because its closest to people, then Luna because there have been people on it, then Sol be the worst because no human has been there?
The naming scheme is too "clever."
Sure—so, is Sol 109.2x better than Terra? Or 1.304x10^6 better?
In the introduction video they say 5.6 Sol autonomously post-trained 5.6 Luna. Curious what this means.
Sounds like they gave it a goal to hit certain benchmarks and just let it have its way with the base Luna model.
Is any of those comparisons about Pro vs non-Pro (Pro is only available in $100+ plans)? I am curious about that but I think Sol, Terra, Luna are different sizes of it without the Pro part, and I want to know how much worse do I have it on the $20 plan compared to if I upgrade.
This marketing video on the page is nice!! can't wait for the hardware to get cheaper to live the AI life i wanna live.
Oh man, I love capitalism spoiling us here. I was just enjoying my extra Fable credits, now I'll switch to using 5.6 this weekend. I was planning to ration my Anthropic credits, I guess now I do not have to. And I was half wondering if exactly this would happen: right when Fable usage credits were starting to kick in for people, OAI swoops in and takes the puck. As much the AI craze is crazy, this play by play part is pretty fun.
top it off with anthropic stressing about the release and resetting usage to 0 for the week just now.
8% on ARC-AGI-3, they actually got some traction going...
note that ARC-AGI-3 has its rules changed.
before today all the contestants were capped at $10k
> Instead of requiring developers to script every step or passing every tool response back through the model, Programmatic Tool Calling in the Responses API can filter large amounts of intermediate data, retain only what matters, and adapt its workflow along the way.
this seems very interesting
I use 5.5 a ton. It's immediately apparent that 5.6 is truly a better model. Hope they don't lobotomize it later.
The claims are pretty bold. I think 5.6 may exceed Fable.
So glad Fable limits just got reset. Thanks OpenAI.
prompts -> loops -> slingshots?
Its an extremely capable model. I think the way we need to approach works shifts again. We need to get our harnesses/workflows to let it gather some momentum on the first couple rounds but then we also need to structure it so that it can slingshot and accomplish the long range goal.
I wish model launches were like proper product releases
it's impossible to _try_ it out on release!
it's not on their codex subscription, or the web/mobile chatgpt interfaces, or aws bedrock, etc. I just cant find a working endpoint with the latest model after they announce
The announcement says they're rolling it out over the next 24 hours or so. I think it's reasonable to do a slow-roll-out release for one of the most used products on the internet.
I wonder what increment of progress will be achieved by the next billion dollars
Is it available in EU? I only see 5.5 still :-(
Wow, the "Agents' Last Exam" graph looks unreal!
I mean the y axis is deceptive to make it seem like greater gains since it starts at 30%, when in reality the differences aren't great.
Even worse, it's not a fair comparison: they purposefully just used "adaptive" instead of "max" for Fable.
What about the graph looked so unreal to you?
That’s because it’s bullshit
Dirac (https://github.com/dirac-run/dirac, https://dirac.run/) now supports gpt-5.6. This thing does now seem to be on the chatGPT/codex accounts yet.
UPDATE: it is now available in chatGPT account also, they rolled it out
Will be there soon according to the last commits in the codex repo: https://github.com/openai/codex/pull/31684/changes
Also, confirmed it works for me by using --model gpt-5.6-sol
There is an issue on the page that causes the benchmark tables to get cut off. If you highlight and drag right you can see a few more models like Gemini and Claude Opus. It's also interesting that they introduced explicit caching, which is something that only Anthropic had for a long time.
GPT‑5.6 system card https://deploymentsafety.openai.com/gpt-5-6/gpt-5-6.pdf
Looks like a great set of models, but there are about 20 different thinking/model levels here in this family and they are very complex to pick the right one for the task
E.g. for GeneBench Pro, it looks like you would always use GPT-5.6 Sol over Terra/Luna, its pareto optimal.
For Agents Last Exam, you would maybe want Luna, then Terra, then Luna, then Sol as you increasingly budget for tasks.
I feel that there may need to be a new auto mode in many of these cases. It selects the best model and thinking given a particular problem.
Feels like it's going to have to go that way eventually, because here we have about 20 different model and thinking levels you could use, and they're not obvious which ones are right for the given use case.
> GPT‑5.6 also introduces more predictable prompt caching, including support for explicit cache breakpoints (opens in a new window) and a 30-minute minimum cache life.
Great to read they are moving away from the 5 minute cache defaults. Hopefully other providers follow soon!
Sounds like a perfect fit for a minimal or bespoke harness?
The cost & output token charts are useful but I wish I could view them more like a 3D surface. Like the CS:APP memory mountain charts.
I wonder how long model size and effort will be a few discrete points instead of continuous.
We have an official pelican on a bicycle from the OpenAI livestream:
https://imgshare.cc/mz9xwut3
holy moly it's in THREE dimensions!
AGI solved
So it's failing epically because it generated a tricycle instead of a bicycle?
The livestream presenter goofed and said the test is typically a tricycle, so I wonder if that's just a coverup.
The site they showed had the pelican riding a number of things, including a bicycle and another pelican (no, not in that way).
"GPT‑5.6 is available starting today across ChatGPT, Codex, and the OpenAI API. The rollout is starting globally now and will continue gradually toward full availability over the next 24 hours."
I am on Plus subscription and see Terra and Luna in Codex, but no sign of Sol. Will it be available only on Pro plans?
I an on Pro and it still returns "The 'gpt-5.6' model is not supported when using Codex with a ChatGPT account"
UPD from announcement: "The rollout is starting globally now and will continue gradually toward full availability over the next 24 hours."
My Codex app got upgraded to the new unified ChatGPT app. I don't see Sol available though. Only Terra and Luna. I'm on the Pro plan. Anyone else see it?
Same, no Sol (i'm on plus)
I find that 5.5 gives me far fewer refusals than Anthropic models for security and reverse engineering work. I hope the same is true for 5.6.
GPT-5.6 Sol, Terra, and Luna. at this rate GPT-6 will be named after a parking lot and GPT-7 after whatever Elon names his next kid.
I wish they had kept their previous sensible naming convention instead of this celestial Sol, Terra, and Luna mumbo-jumbo
So is 5.5-Daybreak still relevant for cyber security give. 5.6 capabilities?
They have a fantastic media team.
If it's not dangerous enough to be classified as WMD by USG, who's interested.
cursor benchmarks with GPT 5.6 in picture, a good reason to stop using opus.
https://cursor.com/evals
The good news you don't have to send your dollars to China to fund ai dictatorship, in russia, north korea, african countries and south america.
Will this run on Cerebas? I'm really looking forward to that.
Sam Altman confirmed during the initial limited release that Sol will run on Cerebras at 750 tok/sec.
>Even at medium reasoning, it beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost.
Sounds great.
Also latency looks very good.
I'm disappointed these models continue to be closed source and so expensive.
Open weight models being 10x or more cheaper is just so much more of an unlock than incremental gains for me.
Like the last time, again they failed to note whether there is an Instant model or when it might become available.
Just a day before my $100 subscription expires, perfect
where is it? Still not accessible...
"GPT‑5.6 is available starting today across ChatGPT, Codex, and the OpenAI API. The rollout is starting globally now and will continue gradually toward full availability over the next 24 hours."
"On Agents’ Last Exam (opens in a new window), an evaluation of long-running professional workflows across 55 fields, GPT‑5.6 Sol sets a new high of 53.6, eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 points. Even at medium reasoning, it beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost. That efficiency extends to smaller models, which are essential to making intelligence more abundant and affordable: GPT‑5.6 Terra and GPT‑5.6 Luna outperform Fable 5 at around one-sixteenth the cost. "
Some pretty big claims and results! Excited to see how it feels during usage.
I use Fable and 5.5 extensively and I still find both have a place in my toolkit, i.e. Fable IS good but it isn't perfect, and it's still better to play them off against each other. I have Fable and 5.5 write plans and have them adversarially review each other's plans.
Having this amount of competition in the coding model space is good for all of us.
Here's me using a Gemini chat log scraper (from Gdrive) then dumping my prompt+Gemini response into local AI
Never go over the free limits in Gemini Pro.
Gemini is great at research and architecture, and my 30 years experience in programming everything; for fun or work; means together there is little to no code slop.
Add to project repo some git submodules of reference source code; boom, bobs your uncle
Zero reason to sign up for OAI or Claude. With employers realizing the costs are more than employees, local models getting more powerful, and models in chips just a few years out, neither of the one note LLM companies without diversified services and R&D portfolios gonna last
If OpenAI can add all the features from CC into Codex i’ll gladly switch.
Which features?
You posted the same comment twice.
> On the Artificial Analysis Coding Agent Index, GPT‑5.6 Sol with max reasoning sets a new state of the art at 80, 2.8 points above Fable 5, while using less than half the output tokens, taking less than half the time, and costing about one-third less.
> That advantage extends across the family: Terra performs just above Fable 5, while Luna outperforms Opus 4.8; each does so in roughly one-third of the time, with about half as many output tokens, and at approximately one-quarter the estimated cost.
Wow. I don't believe it. Every indication and twitter post told me that Fable is much more intelligent than Sol and here we are told that even Terra outperforms Fable?
Not only that, Sol doesn't even come with run time classifiers. So it is even more suspicious.
What's even stranger is that OpenAI is directly referencing a competitor in this direct way.
The meat of the report for SWEs:
SWE-Bench Pro Sol: 64.6% Fable: 80% Opus: 69.2% (!!!!)
So, it still trails Opus, significantly, and is not a next-gen coding model like Mythos/Fable 5.
Disappointing to say the least, but somewhat expected.
SWE-Bench pro is pretty much useless now even though many ppl still look at it. OpenAI published a report yesterday saying so as well. Only look at DeepSWE and FrontierCode right now for coding imo.
Amazing, a company that does poorly in a benchmark says that benchmark is useless...
SWE-bench series just aren't that great by today's standard, even Anthropic previously stated Claude had memorized solutions for the non Pro version of the benchmark, I suspect the recent increase in the score for the Pro version probably also had similar behaviors.
But anyway, I think it's pretty useless to look at SWE Bench's now when other way better benchmarks exist.
You've overstated the conclusion. The SWE-bench series has had issues since its inception.
OpenAI no longer recommends SWE-Bench-Pro as a benchmark: https://openai.com/index/separating-signal-from-noise-coding...
Makes sense why they released an entire study yesterday discrediting SWE-bench Pro.
And they'd be right, it's an almost saturated benchmark where even some subpar open source models score very well on. And most models are clustered within a small range so it really doesn't tell you much.
Do they expect us this model is 15ppt more accurate at half the price of fable? What’s going on?
Not available - checked and it's not there.
As usual, even though GPT-5.6 is releasing today, the rollout in ChatGPT and Codex will be gradual over many hours so that we can make sure service remains stable for everyone (same as our previous launches). We usually start with Pro/Enterprise accounts and then work our way down to Plus. We know it's slightly annoying to have to wait a random amount of time, but we do it this way to keep service maximally stable.
The timescale is typically hours not minutes, so if you don't see it now, I'd try again later today.
We mention it will be a gradual rollout over the next 24 hours in the Availability section at the bottom of the blog but I admit it's pretty buried.
(I work at OpenAI.)
Is this bug fixed with 5.6? If not, it probably doesn’t matter which version Codex users are getting because the overall result is dramatically worse than stated by Open AI advertising: https://github.com/openai/codex/issues/30364
Understood thanks; will 5.6 fix this issue that makes Pro unusable?
https://github.com/openai/codex/issues/30364
"GPT-5.5 Codex reasoning-token clustering at 516/1034/1552 may be leading to degraded performance on complex tasks"
It's available in Cursor now.
on Plus I see Terra and Luna, but not Sol
They have really been stringing us along for the past few weeks.
Thought Fable was great
if OpenAI adds all the features from CC into Codex, i’ll gladly switch.
What features are you missing? That you can't add through skills?
The marketing team must've done research that said "people are starting to think that you guys are evil-water-stealing-lay-off-loving-bubble-bursting scumbags" and decided to really lean into the small family business and happy font vibes!
CTRL-F: Fable
15 hits
Holy shit. They must be feeling very threatened by Fable if they're spending this much energy talking about it in the release notes for their own model.
In the past they received a lot of hate for not comparing to the competition.
yikes - looks like you need to go back to stats school
gemini - 13 hits
opus - 18 hits
So they are more threatened by opus than fable, or are they almost as threatened by gemini as they are by fable?
The second paragraph has four mentions of Fable. I think that makes my case pretty clearly.
Anthropic fumble of Fable's release will go down in the history books, makes sense for OpenAI to run with it.
Apparently it significant outperforms fable on both an intelligence and cost index.
I don’t believe it at all and I don’t think anyone else does either.
I believe that it outperformed it on benchmarks.
Downvoted comment but I did find this comparison aggressive and tacky.
At this point, they are just changing the decimals to stay relevant and in the news.
Anthropic should be grateful OpenAI did not borrow "Epic" and "Legend".
I expect OpenAI names to be "fabulous", "glorious", "empowered", "delicious" etc.
Most importantly, the cost:
> GPT‑5.6 is priced per 1M tokens across three model sizes: Sol is $5 input / $30 output; Terra is $2.50 input / $15 output; and Luna is $1 input / $6 output.
Just as expensive as Fable 5. But of course, another slot machine upgrade but the costs will keep going up and the open weight models from china will continue to race everyone else to $0.
Looking forward to the next version of GLM, Qwen, Deepseek and Minimax.
That's wrong. GPT 5.6 Sol looks to have the same price as GPT 5.5, apart from a new pricing fee for cache writes. Fable 5 is $10 input / $50 output.
Also watching deepseek closely. Seems like US frontier labs only know how to throw money at things as opposed to actually make smart improvements to the algorithms.
To be fair, DeepSeek doubled prices during the peak Chinese workday. (Which admittedly doesn't affect me much.)
The way they talk about cyber security fixes makes clear that they are in bed with the government in order to get ahead of Anthropic.
All of them closely collaborate with the government. LLMs are a national security priority and are vetted. Claude AI was used by Palantir's Maven to target the Minab school that led to a triple tap strike killing over 150 schoolchildren.