SemiAnalysis says ChatGPT Pro yields $14,000 in API value, prompting counterarguments that developer APIs are wildly overpriced

plugging my pinned post about why claude is so egregious as a product here for additional context since this thread is blowing up.

@usr_bin_roygbiv Even if they are subsidized think powerusers on here underestimate how many people fall in the <25% plan utilization part of that table where they make money anyway

Interesting claim from SemiAnalysis.

AI subscriptions are dramatically underpriced versus API usage:

- For heavy coding/chat users, the subscription can be 40–70× cheaper than paying API rates; the API is mainly better when you need automation or product integration.

- a $200/month ChatGPT Pro plan can provide about $14,000/month of API-equivalent usage, while a $200/month Claude Max 20x plan can provide about $8,000/month.

@doodlestein Every decision they have made in the last six months leads me to believe they are very clearly juicing P&L and knowingly cannabilizing customers to justify valuation for an IPO despite not having a competitive product:

@rapidchair oh definitely. vast majority of poeple do not fully utilize their plans even still

@usr_bin_roygbiv bingo. things are subsidized, but not the way most people think.

API cost is inflated, and all the people not hitting limits on subs are just adding the the margin.

why would the companies give it away when they know we'll pay whatever?

@StoicYield yes lmao

@SubleBoom @usr_bin_roygbiv Deepseek is working on efficiency, the whole point of V4 was it wasn’t an incredible model but it was groundbreaking in efficiency.

Western frontier labs are moving the opposite direction, throwing compute and parameters at the problem racing to make the strongest model.

@usr_bin_roygbiv You are very wrong, I have the Anthropic figures from 2024 and even back then they were heavily subsidised.

Your underestimating the size of the models (Opus 3T, Mythos 5-10T) the expensive labour costs, costs of running and purchasing compute.

A single B300 draws 1400w. A populated rack draws 125kw. They have hundreds of these.

+ paying contracts to Space X for compute. $15B per year. + median $420k salary per employee.

And you’re claiming they are profitable on plans that give $8k worth of credits for $200. A factor of 40x.

I can’t explicitly state the figures from 2024, but all I’ll say is costs were an order of magnitude higher than revenue.

And I suspect the figures are worse now as the competition for compute has grown.

@0xdippo use it in omp it is night and day antigravity and gemini cli are hot dogshit see my pinned post

@usr_bin_roygbiv The $20 subscription is probably so profitable because 95% of people don’t do shit

@usr_bin_roygbiv @yacineMTB i don’t think this is a hot take at all. as you already said if you’ve used open weight models you already know this

@Youssofal_ the models are quite literally 100x as token efficient and the hardware is 10x as power efficient

@usr_bin_roygbiv I don't know why people are like "I'm getting near sonnet performance out of 1500 dollars worth of hardware and an open model" and they assume that trillion dollar hyperscalers aren't getting 1000x better bang/buck from their inferencing hardware.

@0xdippo only people saying this shit are people who haven't seen the insane limits on gemini/dsv4

@albustime you’re right it’s even cheaper per user on h200s and b300/

@usr_bin_roygbiv I think also even if they lose money on the max plan it s not 20x especially Anthropic. They probably have all sorts of levers they can pull. Users of a max plan tend to be power users most are gonna use a lot of it. Very dumb business decision to not optimize that shit