Up to 80% decline in a few months: why OpenAI and Anthropic break prices of their LLM

Data scientist Max Woolf traces the price war between OpenAI and Anthropic since June, and attributes it to competition from open models served by dozens of hosting companies.

In a long post, Max Woolf (minimaxir) reconstructs the pricing timeline of the two leaders since June. On the OpenAI side, GPT-5.6 Luna lost 80% of its price in July, then GPT-6 Sol arrived at $2 per million tokens in input and 10 in output, half as much as its predecessor; GPT-6.1 Sol then halved the cost of reading the cache. On the Anthropic side, the launch discount of Claude Sonnet 5 has become permanent, Fable 5.1 has reduced the price of the cache by 75%, and Opus 5.5 costs 40 to 60% less than Opus 5 in total, with the company explaining that it requires less computation to serve.

The cache, a hidden expense item

The author insists on an often overlooked detail: for programming agents, the replay of the cache, which contains all the past conversation, represents 96 to 99% of the tokens invoiced. A 75% drop on this item therefore reduces the real bill by almost as much. The data from Vercel's gateway shows a cost per token that has fallen to 43.3% of its April level, even if total expenses increase with the quality of the models.

For Woolf, these declines are not only explained by the rivalry between the two laboratories. Changing the model is now a matter of simple adjustment, and open models exert a lot of pressure: the latest version of DeepSeek V4.1 Flash is offered by 29 hosting providers on OpenRouter, compared to 13 for the previous one in May, and Kimi K3.1 Pro rivals Opus 5.5 for half the price. He concludes that the thesis of an artificially low price, which is likely to rise once users are captive, does not hold up well in such an open market.