No amount of valuation can fix global supply issues for GPUs for inference unfortunately.
I suspect they're highly oversubscribed, thus the reason why we're seeing them do other things to cut down on inference cost (ie changing their default thinking length).
maybe, but the response to GPU shortages being increased error rates is the concern imo. they could implement queuing or delayed response times. it's been long enough that they've had plenty of time to implement things like this, at least on their web-ui where they have full control. instead it still just errors with no further information.
i notice that as well. most of the time when i see those it has a retry counter also and i can see it trying and failing multiple requests haha. almost never succeeds in producing a response when i see those though, eventually just errors out completely.
Wouldn't that be good? I remember back in the day you could only get Gmail thru an invite, it was an awesome strategy. "Currently closed for applications" creates FOMO. They'd just need to actually get the GPUs in relatively short supply. They could do it in bursts though, right? "Now accepting applications for a short time."
I'm not an internet marketer but that sounds like a win win to me. People feel special, they get extra hype, and the service isn't broken.
Yep, daily haha. Well at least this time they aren't just silently reducing thinking on the server side, which ended up making a mess in my codebase when they did that last time. I'd rather a 500 than a silent rug-pull.
I literally just came to HN to ask if I was alone with the acurséd "API Error: 500 {"type":"error","error":{"type":"api_error","message":"Internal server error"},"request_id":"…"}" greeting me and telling me to get back to using my brain!
500-series errors are server-side, 400 series are client side.
A 500 error is almost never "just you".
( 404 is a client error, because it's the client requesting a file that does not exist, a problem with the client, not the server, who is _obviously_ blameless in the file not existing. )
I know you added the defensive "almost" but if I had a dollar each time I saw a 500 due to the session cookies being sent by the client that made the backend explode - for whatever root cause - well, I would have a fatter wallet.
Indeed, and also there's a special circle of hell reserved for anyone who dares change the interface on a public API, and forgets about client caching leading to invalid requests but only for one or two confused users in particular.
Bonus points if due to the way that invalid requests are rejected, they are filtered out as invalid traffic and don't even show up as a spike in the application error logs.
I know that in principle this is true. However, I have seen claude shadow-throttle my ipv4 address (I am behind CGNAT), in line with their "VPN" policy -- so I do not trust it, frankly.
Funny that I just saw this after have "Console temporarily unavailable". I am currently at the stage that: 1) I think Claude Code is very impressive 2) I think pretty much everything else about them is terrible.
* Support really poor, raised a ticket last week and have heard nothing back at all
* Separation of claude.ai accounts and console accounts is super confusing
* Couldn't log into the platform since I had an old org in the process of deletion even though I was invited to a new one (had to wait 7 days!)
* Payments for more API credits were broken for about a week
* Claude chat has really gone to s*t unless it always was. Just getting back terrible answers to simple questions.
* The desktop app is a web app pretending to be a desktop app that doesn't always know it is a desktop app so you get things like, "this will only work in the desktop app". Yes I know, this is the desktop app! "Oh sorry about that but you need to use the desktop app".
* mcp integration and debugging is dreadful, just a combination of generic "an error ocurred" and sometimes nothing at all
* MCP only supports OAuth for shared connectors but auth key doesn't work even with "local" servers that are not necessarily local, just the config is local.
yeah the desktop app forgets it's the desktop app. claude code feels local right up until the api starts coughing up 500s. same thing, just in a terminal instead of a window.
A few hours ago I noticed a considerable decline in code quality. It seemed the model got downgraded so I switched to codex. Anybody else noticed this? It starts to switch from deep reasoning and trying to fully grasp architectural changes to trying to solve things on a very adhoc basis. Maybe that's just my imagination or maybe that's Anthropic trying to balance the load before being fully overloaded.
If you think AI can replace an SRE in 2026 April, I've got a bridge to sell you. I'm not saying "don't use AI." I'm saying don't turn off your brain and let AI drop your production database.
They compute as total minutes down as a fraction of total time. What this means is that being down, say, 55min during peak-use counts the same as being down 55min when nobody is trying to use it. And congruently it counts being up when nobody is trying to use it as the same as being up when everyone is trying to use it.
Over 90 days though. They had a lot fewer users in February. (And even then, these outage durations seem to add up to more than the error budget 99.26% implies...)
Have anyone found good techniques to get a session out of Claude Code, so that I can point another tool at it and pick up there?
This always seems to happen at the worst possible time, after having spent an hour getting deep into something – half finished edits across files, subagents running, etc.
Honest suggestion - ask the agent to figure a compat shim out, the files are jsonl stored at your ~/.claude/sessions you can most likely just reshape it to work on OpenCode or similar, or have a different Claude Code config that points to OpenRouter or other API style endpoint CC supports and then you can swap accounts and it should still work!
I'm trying that out with Cursor now. But it does take some work to get it to the same state with subagents and making sure it understands the state of the progress that was interupted.
But it seems worth the time to get a solid skill defined up and running that can do this, given that's it's an almost daily event by now.
Maybe a good candidate for a Claude Routine!
"By this time each day, brace for upcoming outage by preparing a comprehensive information package for Cursor to take over your work on active sessions" ...
I don't use any other harness, but I have a cron that picks up changes in my jsonl every X minutes and writes them to a SQLite database with full text search. I also have instructions in my user level claude.md (applies to all projects) to query that database when I'm asking about previous sessions. That's my primary use case where I want it to grab some specific details from a previous session. That's my primary use case. I have terrible context discipline and have built some tools to help me recover from just continuing a different task/conversation with the wrong context.
I could search it myself, but haven't needed to. Getting it out of SQLite into some format Cursor understands should be trivial.
Copy-pasting previous plain-text conversation + a snippet of "inspect the current git changes, and resume where you left of" tends to do the trick, at least in Codex, worked with moving from CC, Gemini and a bunch of others.
It seems like Claude has taken Github's place in terms of developer reaction to it being unavailable. It's like everyone forgot how they did things 18 months ago.
2 days to fix a major issue where we can't login from any sort of web terminal? (even typing manually doesn't work as `n` char auto-exit the frame). All kind of CI/CD pipelines are broken if you were delogged for any reason.
Yea, it's peak time. They don't have enough compute. Why do you think they are banning external subscription use. They sell subscriptions. They don't need people to use CC. That doesn't matter. And yet - they won't have people using their service outside of CC. Something is fishy.
OpenAI is very good in terms of not having as much outages as Anthropic, but almost all products except Codex and the pro model is unimpressive, anthropic has the opposite situation.
for the longest time, anthropic with claude+code was the goat and everything else was mid at best, sounds fmailiar? right now codex is just a pleasure to work with while anthropic is dropping balls left and right, hopefully the planned IPO makes a bit of fire under their asses to get their vibecoded messes sorted and the core experience competitive again. even Opus 17 won't fix this when it gets nerfed or straight up isn't reliable or too expensive for more than 3 prompts a week.
But I thought coding was solved? I guess having a single 9 of availability is something we need true AGI for, we should probably give OpenAI and Anthropic another gazillion dollars to burn through to figure this out!
good advertisement now to shift the tide back to openai that just works and honestly codex with gpt 5.4 is _surprisingly good_ currently, not nerfed or forgetting half the tasks along the way so far. Opus already got worse than sonnet last weeks beyond just crazy token costs, now reliabilty goes to shit and anthropic seems like using it. Meanwhile, delightful of the codex desktop app in fact, stuff seems to "just work" elegantly with good quality.
Same with Reddit. A decade ago it felt like they were down more than they were up. And it didn't slow down their growth trajectory. Instead, as soon as it was back there would be a thousand shitposts about "How did you all survive the outage? Did you <gasp> work?"
I can't speak for all CC users, but I genuinely don't care about the downtime as long as it's resolved in an hour or so. It replaces a manual coding workflow that was also prone to random "downtime" when I got annoyed or had a headache, so it's still a net improvement.
Cross your fingers they're about to drop 4.7. 4.6 came out with a bang, now it seems all the compute bottlenecks just lead to customer frustration as they get closer to releasing next model. Balancing the books over there must be a nightmare, "Well we can piss off every single customer for a week, but we'll be able to release the next model 1 week faster"
Why would anyone assume a new model is dropping when their status page is showing elevated errors? Are they that sloppy that they just let their status systems report failures when they are the ones deploying new infrastructure / models / etc?
I just had to upgrade my plan because I ran out of tokens because medium effort had dementia and things only worked on high. Good to know I'm getting my money's worth...
If you mean Codex is better at planning, I've heard the exact opposite. I'm told it's a beast if you tell it exactly what you need as it will execute it to the T whereas Claude will push back or do its own thing either because it thinks it's wrong or because it's feeling lazy
I've started hitting Codex quota regularly for the first time the last couple of weeks, so I feel like they might be tightening the screws on the $20/month plan too. Someone paying for Max might have to work at it to hit the quota
Well this latest outage has me forming a position that a backup is mandatory. I've been using Codex for adversarial-review, so with this outage I'm now going to ensure the repo is tooled up to use both agents, and when an outage hits just switch over and keep going.
I think the only correct answer here is: It depends, on so many different things. Usage is definitely way more generous with codex and it isn't even close.
You can try out a bunch of models on OpenRouter and see what works for you. Paying per token might be too expensive long term, but definitely a good way to figure out which models you like, and then look at providers.
The other big ones would be OpenAI with Codex and Google with their Gemini and their CLI or Antigravity. Or various IDE plugins or something like OpenCode on the tooling side. GitHub Copilot is pretty cheap and gives you basically unlimited autocomplete and generous monthly quotas that let you try out the most popular models. Also GLM 5.1 is pretty decent if you want to look at other subscriptions. Cerebras Code gave you a lot of tokens but their service wasn’t super stable last I tried and they also don’t give you the latest models.
Personally I just stick with Claude and the 100 USD Max subscription cause it still works really well, even the latest update today to the desktop app made it better (was slow and buggy a month ago, has been gradually getting better) and the Chrome plugin lets me get fully autonomous loops working.
I use both claude code and opencode w/ a fireworks.ai firepass subscription.
Everything I set up in claude code I mirror in opencode.
I do more memory oriented things in CC and I end up doing a lot of things in opencode, especially when I want long-running things and I don't want to be limited by budget.
I was/am a fan of z.ai’s GLM models as a drop in replacement for Claude. But they more than doubled their prices recently. Still a ok alternative, but not really an amazing deal anymore.
does google actually host anthropic models themselves?? surprised anthropic allows that, given how notoriously crazy they are about distillation or weight leaks or any hints of their models being used in the wrong way.
Yes, we host themselves, acting as the data processor which can be important for enterprise customers.
From developer experience, hosting them ourselves allows us to take advantage of our unique infra and deliver fastest time to first tokens of the providers.
- Anthropic introduced stringent limits at peak hours. By "introduced" I mean announced it on a random dev's Xitter account
- Users suddenly started burning through all of their tokens even on trivial tasks. Anthropic never truly acknowledged it, their random devs posted "we're working on it".
- One of the workarounds was to somewhat quietly reduce default reasoning to medium
- OpenClaw and "usage through other tools" banned
- Announce "redesigned Claude Code Desktop App that lets you run many parallel sessions"
- Availability is still circling down the drain
- Dario Amodei is in continuous "trust us we have AGI coding is solved we don't need programmers just give us more money" mode now
Basically pushed the button staying up late finishing something, didn't really factor in a Claude outage in the middle of it, here's to red eyes while I use my clumsy fingers and brain to complete the task the old fashioned way.
While your developers are twiddling their thumbs waiting for Claude to come back online, your competitor is using alternatives to get work done right now and advancing on their go to market timeline.
Claude Code returning: API Error: 500 {"type":"error","error":{"type":"api_error","message":"Internal server error"},"request_id":"---"}
Over and over again!
I suspect they're highly oversubscribed, thus the reason why we're seeing them do other things to cut down on inference cost (ie changing their default thinking length).
Engineer roles dead in 6 months.
You're never gonna guess what software engineers do.
I'm not an internet marketer but that sounds like a win win to me. People feel special, they get extra hype, and the service isn't broken.
In the case of Anthropic is fake availability.
Sam Altman explained the idea is to scale the thing up, and see what happens.
He hadn't claimed to offer a solution to the supply problem that would unfold.
A 500 error is almost never "just you".
( 404 is a client error, because it's the client requesting a file that does not exist, a problem with the client, not the server, who is _obviously_ blameless in the file not existing. )
I know you added the defensive "almost" but if I had a dollar each time I saw a 500 due to the session cookies being sent by the client that made the backend explode - for whatever root cause - well, I would have a fatter wallet.
Bonus points if due to the way that invalid requests are rejected, they are filtered out as invalid traffic and don't even show up as a spike in the application error logs.
8.30am on the US west coast
https://mesmer.tools/random/is-it-peak-hours
* Support really poor, raised a ticket last week and have heard nothing back at all * Separation of claude.ai accounts and console accounts is super confusing * Couldn't log into the platform since I had an old org in the process of deletion even though I was invited to a new one (had to wait 7 days!) * Payments for more API credits were broken for about a week * Claude chat has really gone to s*t unless it always was. Just getting back terrible answers to simple questions. * The desktop app is a web app pretending to be a desktop app that doesn't always know it is a desktop app so you get things like, "this will only work in the desktop app". Yes I know, this is the desktop app! "Oh sorry about that but you need to use the desktop app". * mcp integration and debugging is dreadful, just a combination of generic "an error ocurred" and sometimes nothing at all * MCP only supports OAuth for shared connectors but auth key doesn't work even with "local" servers that are not necessarily local, just the config is local.
You can put those on the health status!
In fact, this proves that there is no AI Bubble and we are massively capacity constrained (aka we under-invested in infrastructure)
Maybe a good candidate for a Claude Routine! "By this time each day, brace for upcoming outage by preparing a comprehensive information package for Cursor to take over your work on active sessions" ...
I could search it myself, but haven't needed to. Getting it out of SQLite into some format Cursor understands should be trivial.
https://github.com/rkuska/carn
'Export the current conversation to a file or clipboard'
> status.claude.com: "All Systems Operational"
Early twitter showed the fail whale as often as it showed tweets and yet it was an unstoppable juggernaut that people kept using.
Android app is still responding but no-go on claude.ai and I can't login with email
status.claude.com has an update:
Investigating - We are seeing increased errors on Claude.ai, API, and Claude Code Apr 15, 2026 - 14:53 UTC
I've never hit the quota on Codex.
On Claude (Code), I used to hit it every other day before switching to Codex.
The other big ones would be OpenAI with Codex and Google with their Gemini and their CLI or Antigravity. Or various IDE plugins or something like OpenCode on the tooling side. GitHub Copilot is pretty cheap and gives you basically unlimited autocomplete and generous monthly quotas that let you try out the most popular models. Also GLM 5.1 is pretty decent if you want to look at other subscriptions. Cerebras Code gave you a lot of tokens but their service wasn’t super stable last I tried and they also don’t give you the latest models.
Personally I just stick with Claude and the 100 USD Max subscription cause it still works really well, even the latest update today to the desktop app made it better (was slow and buggy a month ago, has been gradually getting better) and the Chrome plugin lets me get fully autonomous loops working.
Everything I set up in claude code I mirror in opencode.
I do more memory oriented things in CC and I end up doing a lot of things in opencode, especially when I want long-running things and I don't want to be limited by budget.
- downdetector comment
Caveats: 1) May not be economic for those on flat-rate Anthropic subscription plans 2) I work at Google.
From developer experience, hosting them ourselves allows us to take advantage of our unique infra and deliver fastest time to first tokens of the providers.
- Anthropic introduced stringent limits at peak hours. By "introduced" I mean announced it on a random dev's Xitter account
- Users suddenly started burning through all of their tokens even on trivial tasks. Anthropic never truly acknowledged it, their random devs posted "we're working on it".
- One of the workarounds was to somewhat quietly reduce default reasoning to medium
- OpenClaw and "usage through other tools" banned
- Announce "redesigned Claude Code Desktop App that lets you run many parallel sessions"
- Availability is still circling down the drain
- Dario Amodei is in continuous "trust us we have AGI coding is solved we don't need programmers just give us more money" mode now
[0] https://news.ycombinator.com/item?id=47753710