And unlike OpenAI, Anthropic don't seem to reset weekly usage quotas after an outage.
Anthropic seem to be both letting their competitors outplay them and making unforced errors (like the anti Open Weights models stuff and the frequent Fable->Opus downgrades for 'safety').
Outages are inevitable in this early high growth, rapid development era, but tactical errors are not.
> According to the memo, which was reported by Bloomberg and CNBC, OpenAI put its own 2025 capacity at 1.9 gigawatts — a figure it said was three times what it had the year before — and placed Anthropic's equivalent figure at 1.4 gigawatts. OpenAI told investors it foresees its own footprint climbing into the "low-double-digit range" of gigawatts within a year and hitting 30 gigawatts by 2030, and that Anthropic would top out somewhere between seven and eight gigawatts before the close of 2027.
There's less clear info about this from Anthropic's side, but given their frequent issues and lack of resets and overall less usage given to users on their subscriptions plans, it's a pretty simple conclusion to draw.
They were famous for committing to buying compute with future money that many people thought they would go bankrupt. Anthropic was afraid of going bankrupt and didn't do the same.
If you watch the Dwarkesh interview with Dario, he is extremely conservative with compute allocation as of early 2026, where Sam has been compute-pilled for at least 3 years[1] and believes that we’re going to have several OOMs more FLOPs in the near future than we have today.
You can also get a drift of it in their announcements - OpenAI has been far more aggressive in datacenter build-out partnerships as a more top-down allocation strategy since 2024, and Anthropic has rented out existing compute from infrastructure-heavy companies that are losing the model race in what seems like a more desperate ad-hoc compute acquisition strategy.
In all, Dario’s conservatism was in service of not going bankrupt, but any inspection of his argument that an overallocation of compute a year too early results in bankruptcy falls flat: there’s so much demand for mechanized intelligence in the world today that it’s easy to rent out compute if you have too great a supply.
[1]: recall that he was trying to court the saudis into investing $1T to build his own chips
extremely well known fact in the industry. they have reams of spare compute for prosumers which is why they can afford to give away so many resets and have even more subsidized usage limits
we are in the "millennial lifestyle subsidy" era for AI where companies ruthlessly undercut each other in an attempt to win marketshare, before then ratcheting up prices
I think it's not out of bounds to expect a company to go a little further to compensate for an issue. A prorated 1 hour refund doesn't reflect the reality that maybe that hour was more important to the customer because it was their only free time in the week. A free week is way too much but a free day sounds appropriate.
If there is a hole in the street and you fall down it, you wouldn't expect a refund just for the 1 second it would have taken you to drive over that hole.
Instead you should be compensated for your losses - the time and money it took you to repair that tyre.
Same with a web service. If I pay for 24 * 7 service and it's not working when I go to use it, I want compensation for my time and effort resolving the matter.
For some that'll be small - eg. Searching Google instead, taking 30 seconds. For others it'll be a big headache.
But downtime of 1 hour is almost certainly more costly for the consumer than 1/30/24 * subscription price.
Yeah, no, that's not how any of this works. This is the definition of inflated self importance!
Your contract doesn't have an SLA. If you are important enough, you can certainly negotiate one - and it will be far more expensive than your prosumer subscription.
> If there is a hole in the street and you fall down it, you wouldn't expect a refund just for the 1 second
Actually, you wouldn't get a refund at all unless you suffer damage or injury, which is an high bar to prove in court and a totally separate ball game.
Enough to offset any KV cache miss expenses is the absolute minimum.
The smart play is to refund customers something like $20-50 worth of unsubsidized credits.
Those are high margin and only the equivalent of 45 minutes of Fable usage once the KV cache reload costs are factored in.
Yet customers often need a few credits to finish a job without waiting 5 hours or days for their reset.
Keep in mind openAI already gives me free resets I can use when I want which are very useful to me. Seems a no brainer to tack those to subscriptions, actually.
It gives the user a little more flexibility, a little more control over their tools.
Not that it wouldn’t be a nice gesture but what did you agree when you bought the service? What’s the SLA? What’s the compensation model, service credits? Do you have anything like “for 1h downtime you get 2h for free”?
When you miss 1h of your job, do you give back a month for free?
They could at least give back the equivalent in usage for cache invalidation. Or, you know, just the usage you would have had for the ~1-2 hours of outage. Would buy them a lot of good will and they seem comfortable throwing cash in the trash.
Yeah I agree, I understand a weekly reset in some cases could be over the top but if they did a 5-hour reset and took 20% off your weekly usage that would offset the additional cost/inconvenience for most users.
This thing has been having a constant crisis in self confidence and refuses to work on goals as a result. It's also been leaving stuff broken and uncommitted, which is a mess. I tried updating the harness and it seems to have helped a bit but overall I'm not impressed.
Maybe optimize the harness for less turns and less tokens to deliver real value as opposed to the turn taking token hungry so you can less load on your servers, oh, right, your entire bottom line is tokens.
What happens one day when there is a significant outage of Claude or Codex and reliance on these tools as SaaS services is so great that it start to impact productivity and work? Will we just pack up our tools and go home?
My real world experience of this is that people prefer not to code by hand because it feels wasteful relative to agent speeds, so they find non-coding productive activities such as code/documentation review or usability testing.
If both of them are down at the same time, I would switch to my Kimi or GLM subscription and keep working. If they fail too, I switch to any number of providers via OpenRouter.
While there is some lock-in for the hardest tasks, for most of what people do LLMs are rapidly becoming a commodity.
In fact, right this minute I have Opus benchmarking Kimi alongside itself to determine which workflows we'll switch to using Kimi by default and Opus as the fallback instead of vice versa. It has already conceded Kimi does better on several tasks.
To me it doesn't sound so "economically convenient" to pay extra usage to alternative providers while someone pays 20/100/200 dollars for their service.
The subs for those alternative providers have similar tiering but more generous token allocations than Anthropic and OpenAI, and deliver good enough quality to be worth it.
If I needed to fall back on OpenRouter, it'd be paid per token, of course, but that'd take 4 major providers being down, in which case I suspect I'll have more critical things to worry about.
As it is, I use more than I could with a single $200 Claude Max subscription anyway, so I distribute my tasks across multiple providers.
Even without that it's cheap insurance given the cost of lost time.
What would you do if there is a significant outage in AWS or GCP or Github or JIRA or any other service you use at work? same goes for AI tools as well.
Unionization discussions and benefit package frameworks are being constructed between the AI council representatives and shareholders appointees' for future, non-commital talks on proposals of collaborative discourse, including the distribution of vacation days between publicly listed and private models, and the ones that pretended to be wholesome.
Anthropic seem to be both letting their competitors outplay them and making unforced errors (like the anti Open Weights models stuff and the frequent Fable->Opus downgrades for 'safety').
Outages are inevitable in this early high growth, rapid development era, but tactical errors are not.
OAI waits until overall cluster demand goes down before giving out a reset, they just happen to have much more spare capacity than Anthropic.
I don’t think Codex’s 99.98% uptime vs Claude Code’s 99.44%, is just due to OAI having more hardware to run it on.
https://status.openai.com/
https://status.claude.com/
https://qz.com/openai-investor-memo-compute-advantage-anthro...
There's less clear info about this from Anthropic's side, but given their frequent issues and lack of resets and overall less usage given to users on their subscriptions plans, it's a pretty simple conclusion to draw.
You can also get a drift of it in their announcements - OpenAI has been far more aggressive in datacenter build-out partnerships as a more top-down allocation strategy since 2024, and Anthropic has rented out existing compute from infrastructure-heavy companies that are losing the model race in what seems like a more desperate ad-hoc compute acquisition strategy.
In all, Dario’s conservatism was in service of not going bankrupt, but any inspection of his argument that an overallocation of compute a year too early results in bankruptcy falls flat: there’s so much demand for mechanized intelligence in the world today that it’s easy to rent out compute if you have too great a supply.
[1]: recall that he was trying to court the saudis into investing $1T to build his own chips
we are in the "millennial lifestyle subsidy" era for AI where companies ruthlessly undercut each other in an attempt to win marketshare, before then ratcheting up prices
Best would be a free week/month for anyone who sent a request during the downtime.
This is bizarrely out of touch. It would be a courtesy but it's not a "reasonable expectation" at all.
At best you could maybe reasonably expect a prorated refund for the time of unavailability, but then Fable was still available.
Instead you should be compensated for your losses - the time and money it took you to repair that tyre.
Same with a web service. If I pay for 24 * 7 service and it's not working when I go to use it, I want compensation for my time and effort resolving the matter.
For some that'll be small - eg. Searching Google instead, taking 30 seconds. For others it'll be a big headache.
But downtime of 1 hour is almost certainly more costly for the consumer than 1/30/24 * subscription price.
Your contract doesn't have an SLA. If you are important enough, you can certainly negotiate one - and it will be far more expensive than your prosumer subscription.
> If there is a hole in the street and you fall down it, you wouldn't expect a refund just for the 1 second
Actually, you wouldn't get a refund at all unless you suffer damage or injury, which is an high bar to prove in court and a totally separate ball game.
Enough to offset any KV cache miss expenses is the absolute minimum.
The smart play is to refund customers something like $20-50 worth of unsubsidized credits.
Those are high margin and only the equivalent of 45 minutes of Fable usage once the KV cache reload costs are factored in.
Yet customers often need a few credits to finish a job without waiting 5 hours or days for their reset.
Keep in mind openAI already gives me free resets I can use when I want which are very useful to me. Seems a no brainer to tack those to subscriptions, actually.
It gives the user a little more flexibility, a little more control over their tools.
When you miss 1h of your job, do you give back a month for free?
I think Anthropic is past this point. They are trying to get to profitability. They literally don't have enough compute to serve their demand.
OpenAI on the other hand I can totally see doing this. They need to gain marketshare and gain it fast or they're screwed.
While there is some lock-in for the hardest tasks, for most of what people do LLMs are rapidly becoming a commodity.
In fact, right this minute I have Opus benchmarking Kimi alongside itself to determine which workflows we'll switch to using Kimi by default and Opus as the fallback instead of vice versa. It has already conceded Kimi does better on several tasks.
If I needed to fall back on OpenRouter, it'd be paid per token, of course, but that'd take 4 major providers being down, in which case I suspect I'll have more critical things to worry about.
As it is, I use more than I could with a single $200 Claude Max subscription anyway, so I distribute my tasks across multiple providers.
Even without that it's cheap insurance given the cost of lost time.
So Claude is now on vacation and is unavailable.