This model was likely trained months before deepseek released their paper.

Doesn't mean they didn't apply something similar. They could have also come up independently with their own version, the speculation is not they copied it, rather that they have performance breakthroughs which perhaps is a result of work in same domain

Models are generally posttrained to a window shorter than you think.

I suspect the only reason those alternative providers have better up-time and more generous quotas, is because they don't have nearly the same amount of demand. Notice that Deepseek recently had to increase their pricing, once it gained it popularity.

Demand > supply. It's impressive that customers have not migrated en masse to other providers, given the frequency of these outages. Perhaps switching costs are greater than some would believe. Or, qualitative differences between models continue to exist, despite matching on public benchmarks.

I think a lot of people, like myself, have. Codex paid users have grown a lot in the past few months.

check their status page...

That's quite a reach. More likely someone merged and deployed their vibe-coded PR and is now figuring out how to bring the service back up.

If there's a lot of demand for rollercoaster rides, the rollercoaster will not stop operating; instead the queue of people in front of it will increase.

It's not like a bridge or elevator where we have a certain number of people that can use it, and if one more person joins, the whole structure breaks apart and everybody perishes.

Those guys are running a website that provides an interface for some specific hardware. Just like file hosting providers back in the day selling their terabyte-sized hard disks in 100MB-increments.

>> Perhaps switching costs are greater than some would believe.

I haven't switched because there's nothing to switch to that is anywhere as good. I've been making dedicated attempts at using Sol but it falls short, despite what some people claim.

I did migrate 80% of my tokens. But for some tasks claude models are still the best.

Easy workaround is to work outside US peak hours (europe morning). I love this outages, I am hardly affected, and weekly reset usually promptly follows!

Demand is actually shared between users and internal. Anthropic uses their own compute for training and building.

If a meaningful chunk of customers moves to some other service, that service will get swamped and fall down.

If only every outage was due to “demand”.

Reduced weekly limits, higher prices than competitors, paying well above spot price for space-x compute, all point to supply issues.

Do you really think it’s a coincidence claude outages always happen when US and EU workdays overlap?

Anthropic is a trillion dollar company and employs way more skilled high-paid engineers than you are btw, you really think it’s a systems issue not capacity

Translation, moderation, classification, guardrails, etc..

This is brutal, especially for small transactions. Note each failed attempt is another screening fee. Unfortunately, there is no good competition.

Not just each failed attempt, but every transaction. Previously, Radar only checked the card on a subscription at first use. It will now check every transaction unless you opt-out. So not only do you have to opt-out of Radar Standard, you also have to opt-out of having every subscription transaction run through Radar.

Demand > supply

Same

By what measure is there an overbuild? Every metric I look at, shows inference unable to satisfy current demands.

I know companies paying for AI and not training people to use it, so spending is higher than usage on those.

20k is small potatoes for the marketing impact.

They are doing both. Distilling Mythos down to affordable models, so they can continue to fund the business. And training Mythos level models at the high-end, to expand the frontier.