Florian Kronawitter argues Anthropic must cut prices as Claude 4.7 Opus evaluation costs far exceed GPT-5.5 and Gemini 3.5

Old Russian joke. Son: "daddy, they said that vodka prices are rising. So are you going to be drinking less?" Dad: "niet son, but! you'll have to eat less" (for those in the back row: the dad is your employer, vodka is Anthropic tokens)

@fkronawitter1 I disagree…that is saying that IPhone is too expensive, so Apple should make them less expensive or prepare to lose customers? When you build a better mouse trap, the world will beat a path to your door

@HumbleTrader77 see my previous tweets, many corporates now in pain on token cost and no immediate ROI

@fkronawitter1 @LONGCONVEXITY Perhaps. But the issue is really the other way around. The rest of the pack is either not competitive, too cheap or both for long-term survival. The customer subsidy phase has to end, one way or another, until these chips eat less electricity. That’s unlikely happening short-term

@hervecuviliez @fkronawitter1 @pathway_com You mean this? https://pathway.com/research/beyond-transformers-sudoku-bench

Couple of $, infinitesimal.

@fkronawitter1 Now show the rest of the graph: https://artificialanalysis.ai/?intelligence-efficiency=intelligence-efficiency-cost-breakdown#intelligence-efficiency-tabs

@fkronawitter1 They need to copy some of these open source models with regards to attention.

No reason to reread every token for every new token. Ridiculous and outdated.

https://www.jakecuth.com/work/deepseek-v4-lab/

@fkronawitter1 The others aren't cheaper. They're just losing money faster.

@fkronawitter1 i think google bleeds the other us frontier dry, but it is hard to keep up with the chinese price points

@fkronawitter1 that chart misses prompt caching. once you hash the system prompt across repeated calls, anthropic's input bill drops by maybe half on workloads that reuse a lot of context. doesn't move opus into gemini territory but the raw-rate gap mostly closes for our triage agent.

@fkronawitter1 It’s the equivalent of saying “a Mercedes is too expensive compared to Toyota”

@fkronawitter1 They can’t cut much

@fkronawitter1 Wait for new gen architecture like @pathway_com and you will see the cost going dramatically down with better performances! @zuzanna_pathway can you share the cost to crack extreme Sudoku as a exemple ?

@fkronawitter1 but hear me out there grew there revenue faster than any company in the world….

@fkronawitter1 Lower prices won't solve the problem of old organisations.

If they think they get $20M of savings from $200M in token costs, cutting token price in half brings nothing.

@garybasin @fkronawitter1 They have the highest margins of anyone, and now have the cheapest cost-of-capital of anyone in the business (highest valuation).

Seems like they’re best positioned to cut out of anyone in the industry by far?

Chart is showing max reasoning mode, not how most teams actually deploy it.

Also: Anthropic already cut Opus pricing 67% this year ($15→$5 input). The direction is clearly down.

Add prompt caching (90% off) + batch discounts (50% off) and the effective cost looks very different from this chart.

@fkronawitter1 @grok add more unrelated key colors and indexes

@fkronawitter1 My strategy plan!

⬇️

@fkronawitter1 Hidden costs not taken into account:

support load, failed generations, retries, context bloat, orchestration, eval labor, security review, incident response, vendor lock-in, and the human cleanup tax.

Token price is not system cost.

It is just the receipt they let you see..🛠️