Obviously this is way worse than API overall. However, explicitly nerfing subscriptions leads to huge public backlash, and the rapidly falling cost of intelligence means you'll be able to profitably serve Opus 4.8 level models for $20/month in the near future. We therefore think it's far more likely the labs will withhold new features/models from subscription plans. It will be interesting to see if Mythos ends up being API only. (4/4)

The margin on a subscription plan is a function of the average utilization. If we assume both companies have 75% API gross margins, this results in the following subscription margins. (3/4)

Recently, we purchased one of each Anthropic/OpenAI subscription plan and randomly ran long horizon coding tasks until we exhausted the weekly limit. It's widely believed that a $200/month plan maxes out at ~$2000/month worth of tokens (assuming API pricing). However, we found that the subscriptions are actually far more generous. (2/4)

@SemiAnalysis_ This tracks with my last 3 months

@SemiAnalysis_ so what youre saying is when theyre done using us for training, all the normies will be rugged out of the system

@SemiAnalysis_ However the point of subscriptions is that MOST users don't use the quota they paid for. It's like buffets where heavy eaters are subsidized by 99% of the population not eating their fair share.

@SemiAnalysis_ hey check this out

@SemiAnalysis_ Don’t publicise this so much, there might be a chance that once everyone is heavily dependent on AI then @OpenAI and @AnthropicAI will increase prices heavily/limit the usage to actual billing plan

@SemiAnalysis_ spend $200 for a subscription that gives you $8,000 in inference but provides $10 in actual real world value

In my own testing, the 20x plan gives me precisely $2600/weekly GPT 5.5-API equivalent quota. That value was checked for a lot of usage over multiple weeks, calculated from Codex thread metadata (input costs cached/uncached, output), including when I was on the 5x plan (when it had 2x promo, so 10x) and got $1300/week.

So I think it's closer to $10400/month, still huge but less than you implied :)

@SemiAnalysis_ you get 40x -70x leverage on capital with the max subscription plans. at the cost of rate limiting annoyance.

i think this is a good arbitrage for now. they can decide to discontinue the sub plans anytime tbh

@SemiAnalysis_ it's already announced that fable will be in subs until june 22nd, then it becomes only available on credits which happen to be 30% off today but won't be then i'm sure.

You could say that, but at same time API prices are likely significantly inflated and many users of sub plans probably not using it all too. In the end the math is working out for these companies. And doubt it will change much in future, at least regarding models of similiar scales as opus 4.8/gpt 5.5 - if we get some 10T or 20T models those might not make sense to even sell as part of sub anymore. But don't think most users even need those anyway.

@SemiAnalysis_ How did you account for cache hits in the spend calculation?

@FateOfMuffins @SemiAnalysis_ Yes but most of the 100$ and 200$ subs are definitely power users who max out their limits

@SemiAnalysis_ Every enterprise client forced to pay API prices is crying seeing this right now

@SemiAnalysis_ Generated productivity gains: 17 racist ai pictures against jews, muslims and Indians. 38 pictures of Chinese AI girls (age questionable) showing feet and $8 real world value

@ai_hyperbull @SemiAnalysis_ I am planning to build, plan everything I need in the next 11 days to 22nd. 5 hour alarms, max maxing, so I can avoid paying for the API! Can then spend six months implementing with Deepseek lol

@SemiAnalysis_ Curious if you’re aware of any multiplied usage limits or resets being part of the period you tested

@kimmonismus That's not true any more.

I've used Claude Max before the rate cuts and capped at $5k (constantly paused). Now it's less than $4k in Claude and $2.5k in codex (yes... way less).

If anybody can show me a better gpt stat I'd be impressed and know I'm doing something wrong there.