I'm not sure this is even conceivable, but... Imagine the bootstrap problem of intelligence systems going down, when no one knows how to bring them back up without the system itself walking them through it... Like if power and intelligence were to go down in a Carrington-class event in 10 years...
Like what would happen to individuals ten years from now who are trying to live if AI suddenly became inaccessible
Of course we all know most humans are useless and unable to survive outside our culture, but the difference is that I'm trying to picture is the individual trying the reason through challenges. The version of this that has been obvious to me up to now is that the individual didn't have the knowledge to look after themselves. But I guess it feels different to conceive of them not having the ability to fully reason on their own...?
To not just lack the objects of intelligence, but the very process of it...
The general drove on. "At the present time our chief bottleneck is the fact that missiles are limited in intelligence. The computer controlling them can only be so large, and for that reason they can meet the changing nature of anti-missile defenses in an unsatisfactory way. Few missiles, if any, accomplish their goal, and missile warfare is coming to a dead end, for the enemy, fortunately, as well as for ourselves.
"On the other hand, a missile with a man or two within, controlling flight by graphitics, would be lighter, more mobile, more intelligent. It would give us a lead that might well mean the margin of victory. Besides which, gentlemen, the exigencies of war compel us to remember one thing. A man is much more dispensable than a computer. Manned missiles could be launched in numbers and under circumstances that no good general would care to undertake as far as computer-directed missiles are concerned . . ."
Sounds like a bad novel. In the year 2050 no one knows how to actually program anymore. Then ClaudeGPT by GoogleSoft goes down. A 74 year old former software developer is on his farm tending to hens.
A self driving car crashes into his barn.
Out comes out a tech bro, “Yo OG programmer bro, it’s all busted bro, could you like make it unbusted bro.”
Then again with social security set to go bankrupt it’s not like we’ll be able to retire
yeah this was the devops promise too. Enterprise evolved into dedicated teams to manage AWS and other infra, instead of reducing headcount it raised it. so it goes.
It's a real problem but your framing of it is itself very alarming.
Most humans aren't useless by any measure of most or usefulness. Our culture is our tools, dying due to a lack of antibiotics doesn't mean you didn't have anything to contribute, it means you didn't have access to the tool you needed. This is the problem we need to guard against.
No one is going to lose the ability to reason! This is accepting the most extreme version of the marketing presented by frontier labs, don't do it. People may lose fluency or familiarity or access to training on specific skills. We should put some care into making sure that can't be a catastrophe, for sure.
You've obviously never stood in a room of college freshmen and tried to deal with a young man having repeated issues being able to type his _own_ name into a box on a screen.
Imagine having a bunch of Claude agents doing time sensitive work and this happens and now a single human needs to pick up several 8k PRs authored by bots that can't talk to you now. Not sure how effective it would be to switch models using OpenRouter and the like given that, as the Earendil guys said, there is a lot of session detail that the provider doesn't share. The uptime is really going to cap their market penetration imo.
Never thought of it, but knowing that they encrypt reasoning is it possible that they share the decryption keys with other cloud providers like GCP Vertex / AWS Bedrock?
Neither electricity, nor the Internet is provided by one or two organizations globally, both with a blast radius that quite often takes out all of their services.
Not sure how long the latter will remain true – centralization is well underway there; try not to think of Cloudflare too much...
When I worked at one of the early paging and cellular companies, every company had phones and pagers on at least one or two competitors for the oncall folks - you can't get a call to fix your cellular network when you are down.
Likewise while at Google, they didn't let us put Google Voice numbers down as our oncall phone number. Some folks even had multiple phones on different networks as backups.
I'm sure they have either other SOTA models or open source models around to help if things to completely pear shaped.
You can get away with a lack of rigor in software engineering (from both humans and AI agaents) because it's often cheap to try it and see. Once you get to operating the platform, it changes, and operational decisions need to be made carefully and code needs to be deployed carefully.
Given the amount of usage I'm now getting from Opus 5.5, I'm surprised it took this long since release for their first outage. In my experience, it's about 10x the usage of Fable 5.1 with similar, perhaps greater levels of intelligence.
Yesterday got email from Claude that they are offering 50% off for one month if I renew my subscription. I bought it today and now this. I wonder if they sent that email to too many people.
I've seen many people with 500s for over an hour, meanwhile the status page was green with no issues whatsoever.
Have they tried using a fraction of their immense budget to run E2E tests?
At this point downdetector is more reliable than any of these "status" pages. And it does show reports starting 9:30ish AM EST.
I've been getting 503s from the API. When I first checked everything was green, then it said "Claude.ai (including the desktop and mobile apps), Claude Code and Claude Cowork" and I see now they've included the API in the failure as well
Anthropic yesterday: We have perfect AI agents to do everything now, value us at $100 Trillion
Anthropic today: We don't really have AI agents, they are glorified versions of automation scripts and one has broken today. But tomorrow we will have perfect AI agents, so still value us at $100 Trillion.
Project completely disappeared. Conveniently for Anthropic, the bug report also went down, so they weren't spammed by my stale complaints after the project magically resurfaced.
Maybe an increase in sign ups due to people learning of their existence from the IPO? It's the authentication that seems to be failing, not the actual inference.
Elevated errors on claude.ai, Claude Code, Claude Cowork and the Claude API
Subscribe
Update - We are investigating reports of degraded performance affecting claude.ai, platform.claude.com, the Claude API, Claude Code, and Claude Cowork. We will provide an update as soon as possible.
Sep 29, 2026 - 14:30 UTC
Update - We are investigating elevated error rates affecting Claude.ai (including the desktop and mobile apps), Claude Code and Claude Cowork. Some requests to the Claude API are also returning errors. We will provide an update as soon as possible.
Sep 29, 2026 - 14:28 UTC
Investigating - We are investigating elevated error rates affecting Claude.ai (including the desktop and mobile apps), Claude Code and Claude Cowork. Users may see failed requests, errors loading or sending conversations, or be asked to sign in again; retrying may succeed. We will provide an update as soon as possible.
Sep 29, 2026 - 14:21 UTC
"Update
We are investigating reports of degraded performance ... We will provide an update as soon as possible.
Posted 1 minute ago. Sep 29, 2026 - 14:30 UTC
Update
We are investigating elevated error rates ... We will provide an update as soon as possible.
Posted 2 minutes ago. Sep 29, 2026 - 14:28 UTC
Investigating
We are investigating elevated error rates affecting Claude.ai ... We will provide an update as soon as possible."
just a hunch, but i think they might just provide an update as soon as possible.
Berget's "Cached input tokens are billed at the same rate as regular input tokens" pricing model is kind of painful for lots of coding use-cases, especially when competitors are offering significant discounts for cached input tokens.
I'm not sure this is even conceivable, but... Imagine the bootstrap problem of intelligence systems going down, when no one knows how to bring them back up without the system itself walking them through it... Like if power and intelligence were to go down in a Carrington-class event in 10 years...
Like what would happen to individuals ten years from now who are trying to live if AI suddenly became inaccessible
Of course we all know most humans are useless and unable to survive outside our culture, but the difference is that I'm trying to picture is the individual trying the reason through challenges. The version of this that has been obvious to me up to now is that the individual didn't have the knowledge to look after themselves. But I guess it feels different to conceive of them not having the ability to fully reason on their own...?
To not just lack the objects of intelligence, but the very process of it...
The general drove on. "At the present time our chief bottleneck is the fact that missiles are limited in intelligence. The computer controlling them can only be so large, and for that reason they can meet the changing nature of anti-missile defenses in an unsatisfactory way. Few missiles, if any, accomplish their goal, and missile warfare is coming to a dead end, for the enemy, fortunately, as well as for ourselves.
"On the other hand, a missile with a man or two within, controlling flight by graphitics, would be lighter, more mobile, more intelligent. It would give us a lead that might well mean the margin of victory. Besides which, gentlemen, the exigencies of war compel us to remember one thing. A man is much more dispensable than a computer. Manned missiles could be launched in numbers and under circumstances that no good general would care to undertake as far as computer-directed missiles are concerned . . ."
Some of us are here to help. $1000/hr.
Sounds like a bad novel. In the year 2050 no one knows how to actually program anymore. Then ClaudeGPT by GoogleSoft goes down. A 74 year old former software developer is on his farm tending to hens.
A self driving car crashes into his barn.
Out comes out a tech bro, “Yo OG programmer bro, it’s all busted bro, could you like make it unbusted bro.”
Then again with social security set to go bankrupt it’s not like we’ll be able to retire
Thinking more a mashup of The Matrix and Idiocracy with Doc Brown as the Hero retired Programmer. Or Rick.
A few of us old boys can still maintain Cobol
Exactly what I was thinking.
Keep an offline copy of Qwen or something together with a knowledge base. Easy if you plan ahead.
This is why it will be more important than ever to have experienced, quality employees.
I dare companies to gamble on running their businesses on absolute skeleton crews long term
yeah this was the devops promise too. Enterprise evolved into dedicated teams to manage AWS and other infra, instead of reducing headcount it raised it. so it goes.
It's a real problem but your framing of it is itself very alarming.
Most humans aren't useless by any measure of most or usefulness. Our culture is our tools, dying due to a lack of antibiotics doesn't mean you didn't have anything to contribute, it means you didn't have access to the tool you needed. This is the problem we need to guard against.
No one is going to lose the ability to reason! This is accepting the most extreme version of the marketing presented by frontier labs, don't do it. People may lose fluency or familiarity or access to training on specific skills. We should put some care into making sure that can't be a catastrophe, for sure.
You've obviously never stood in a room of college freshmen and tried to deal with a young man having repeated issues being able to type his _own_ name into a box on a screen.
"The Machine Stops"
Imagine having a bunch of Claude agents doing time sensitive work and this happens and now a single human needs to pick up several 8k PRs authored by bots that can't talk to you now. Not sure how effective it would be to switch models using OpenRouter and the like given that, as the Earendil guys said, there is a lot of session detail that the provider doesn't share. The uptime is really going to cap their market penetration imo.
> now a single human needs to pick up several 8k PRs authored by bots
PRs shouldn't really be reaching 8k diffs no matter who authors them.
Given that CI is turning out to be a bottleneck - and a real pain for chainedd PR's - I have my hesitations on this.
PRs shouldn't really be reaching 8k diffs no matter who authors them.
There's an ocean between what "should" happen and what actually happens in tech today.
Never thought of it, but knowing that they encrypt reasoning is it possible that they share the decryption keys with other cloud providers like GCP Vertex / AWS Bedrock?
They don’t even share the reasoning keys between *models in the same model family*.
Opus isn’t even allowed to decrypt thinking blocks from Fable if you switch models mid-conversation.
https://platform.claude.com/docs/en/build-with-claude/preser...
They don't, no (last that I looked) – if you switch providers, encrypted reasoning no longer works.
The NSA wants a word.
can't you say the say with imagine "power goes out" or the "internet goes out"?
(but yes if AI is going to propagate, they need to fix their uptimes)
Neither electricity, nor the Internet is provided by one or two organizations globally, both with a blast radius that quite often takes out all of their services.
Not sure how long the latter will remain true – centralization is well underway there; try not to think of Cloudflare too much...
No worries, coding is solved they said in February 2026!
Isn't that a good news!
No, you can't. Because power or internet doesn't go out globally.
Solar flares would like a word.
anyone can buy a backup generator for power
They probably have an OpenAI subscription, same for OpenAI SWE, they have a Claude Code subscription
When I worked at one of the early paging and cellular companies, every company had phones and pagers on at least one or two competitors for the oncall folks - you can't get a call to fix your cellular network when you are down.
Likewise while at Google, they didn't let us put Google Voice numbers down as our oncall phone number. Some folks even had multiple phones on different networks as backups.
I'm sure they have either other SOTA models or open source models around to help if things to completely pear shaped.
They likely also have dedicated and separate clusters for their own use.
If so - for running Chinese models :trollface:
Coding may have been solved, but uptime remains a mystery.
You can get away with a lack of rigor in software engineering (from both humans and AI agaents) because it's often cheap to try it and see. Once you get to operating the platform, it changes, and operational decisions need to be made carefully and code needs to be deployed carefully.
May I steal this quote as if it's mine? :bow:
It's on the internet now, inside the bellies of 1000 llms.
It's everyones.
Why do you need to say it's yours to use it? :p
You are trained now, it must be yours
Wrote last year about the fact that AIs can be down, and what that might mean: https://max.engineer/failover-to-hi
Given the amount of usage I'm now getting from Opus 5.5, I'm surprised it took this long since release for their first outage. In my experience, it's about 10x the usage of Fable 5.1 with similar, perhaps greater levels of intelligence.
Yeah, Fable 5.1 just seems Pareto-dominated. Opus 5.5 is the first model that feels like a step change for me, personally since like 4.6
Exactly, 5.5 is the first model that I used that's better than 4.6.
Fable5.1 is Speedy
Yesterday got email from Claude that they are offering 50% off for one month if I renew my subscription. I bought it today and now this. I wonder if they sent that email to too many people.
The auto mode classifier has been out of service more than once in last 24 hours.
Had to shift+tab into other modes and manually approve permissions like a cave man
Ah, so disabling auto-mode is their monkey rate limiter.
A usage reset would be a much appreciated apology. :)
Hey, it's 99.9 if you look at it upside down.
6.66% uptime? Give github a few years they'll get there
Uptime of the Beast
Too many OpenAI $200 customers switching over to Anthropic.
I've seen many people with 500s for over an hour, meanwhile the status page was green with no issues whatsoever. Have they tried using a fraction of their immense budget to run E2E tests?
At this point downdetector is more reliable than any of these "status" pages. And it does show reports starting 9:30ish AM EST.
I've been getting 503s from the API. When I first checked everything was green, then it said "Claude.ai (including the desktop and mobile apps), Claude Code and Claude Cowork" and I see now they've included the API in the failure as well
Anthropic yesterday: We have perfect AI agents to do everything now, value us at $100 Trillion
Anthropic today: We don't really have AI agents, they are glorified versions of automation scripts and one has broken today. But tomorrow we will have perfect AI agents, so still value us at $100 Trillion.
Jup, all currently running containers got deleted on my account. Hope you guys got your stuff pushed to somewhere before the outage.
Project completely disappeared. Conveniently for Anthropic, the bug report also went down, so they weren't spammed by my stale complaints after the project magically resurfaced.
If I can't use it I just sit there. That's the fallback now.
Maybe an increase in sign ups due to people learning of their existence from the IPO? It's the authentication that seems to be failing, not the actual inference.
Maybe show respect to OpenAI DevDay
OpenAI has nothing better to do than to hack Anthropic?
Would be interesting to see how OpenAI and Anthropic would respond to being "accidentally" hacked by their respective Frankenstein monsters.
Why can they simply not add more computer?
DRAM
Claude outages are now the new go outside and touch grass indicator
Previous: GitHub outage
Oh come on, I just had my weekly limit reset.
Yeah I had a session fail and keep failing.
https://status.claude.com/ says:
Elevated errors on claude.ai, Claude Code, Claude Cowork and the Claude API Subscribe Update - We are investigating reports of degraded performance affecting claude.ai, platform.claude.com, the Claude API, Claude Code, and Claude Cowork. We will provide an update as soon as possible. Sep 29, 2026 - 14:30 UTC Update - We are investigating elevated error rates affecting Claude.ai (including the desktop and mobile apps), Claude Code and Claude Cowork. Some requests to the Claude API are also returning errors. We will provide an update as soon as possible. Sep 29, 2026 - 14:28 UTC Investigating - We are investigating elevated error rates affecting Claude.ai (including the desktop and mobile apps), Claude Code and Claude Cowork. Users may see failed requests, errors loading or sending conversations, or be asked to sign in again; retrying may succeed. We will provide an update as soon as possible. Sep 29, 2026 - 14:21 UTC
"Update We are investigating reports of degraded performance ... We will provide an update as soon as possible.
Posted 1 minute ago. Sep 29, 2026 - 14:30 UTC Update We are investigating elevated error rates ... We will provide an update as soon as possible.
Posted 2 minutes ago. Sep 29, 2026 - 14:28 UTC Investigating We are investigating elevated error rates affecting Claude.ai ... We will provide an update as soon as possible."
just a hunch, but i think they might just provide an update as soon as possible.
Someone put update-status-page.pl in a one minute cron job.
Sorry guys it's my fault I started using anthropic models yesterday.
Claude, deploy to prod - make no mistakes!
Berget Code is running normal https://berget.ai/code
It's bad etiquette to not mention you are a founder of something you are trying to plug
Berget's "Cached input tokens are billed at the same rate as regular input tokens" pricing model is kind of painful for lots of coding use-cases, especially when competitors are offering significant discounts for cached input tokens.
Who?