Data Scientist Max Woolf Unpacks Four Months of OpenAI and Anthropic Price Cuts: What Really Made Coding Agents Cheaper Is the Cache Read Price, 96%–99% of Token Spend

On October 25, former BuzzFeed senior data scientist Max Woolf laid out OpenAI's and Anthropic's model pricing since June on his blog: from the 80% cut to GPT-5.6 Luna in July, to Fable 5.1 lowering cache reads from $1 to $0.25 per million tokens, to Opus 5.5 and GPT-6 Sol both cutting prices on September 22, to GPT-6.1 Sol halving the cache read price. Because a coding agent re-reads its entire history every turn, he writes, cache reads make up 96%–99% of total token spend, so a 75% cache read cut is roughly a 70%–75% cut in the cost of use. He attributes the cuts mainly to competitive pressure from open-weight models and the many inference providers on OpenRouter: DeepSeek V4.1 Flash 1017 is now hosted by 29 of them. The post is a personal analysis, and some of its judgments are speculative.

A four-month price timeline

Max Woolf has worked as a data scientist for nine years, was formerly a senior data scientist at BuzzFeed, and maintains open source AI/ML projects. His long post on October 25 starts from one question: why did OpenAI and Anthropic cut their model prices by up to 80% within a few months? The prices below follow his format of input / output / cache read, in dollars per million tokens.

The key steps in his timeline: on July 30, OpenAI cut GPT-5.6 Luna by 80% to 0.20 / 1.20 / 0.02; on September 1, Anthropic released Fable 5.1 at 10 / 50 / 0.25, where the only change from June's Fable 5 was cache reads dropping from 1 to 0.25; on September 22, OpenAI released GPT-6 Sol (2 / 10 / 0.20) and Anthropic released Opus 5.5 (4 / 20 / 0.20) the same day, stating that it requires less compute to serve than Opus 5 and that its pricing reflects that; a week later, GPT-6.1 Sol halved cache reads again to 0.10. Woolf cites Vercel AI Gateway data showing cost per token at 43.3% of its April 2026 level, while aggregate spend keeps rising because models are getting better faster than they are getting cheaper.

The column nobody reads on the bill

The most useful part of the post is a piece of arithmetic. A coding agent carries the whole previous conversation into every turn, and that part goes through the prompt cache; Woolf estimates cache reads at 96%–99% of total token spend, with cache hit rates usually between 90% and 100%. So Fable 5.1's 75% cut to cache reads lowered the real cost of using it by roughly 70%–75%, even though input and output prices did not move at all. Put differently, comparing input and output rates to decide which model is cheaper compares the wrong column for agent workloads.

Why prices fell this way

Woolf rejects the idea that prices are kept low to hook users and will rise later. His reasoning is that models are close to commodities: switching models takes a toggle in tools like OpenCode or Zed, or a different endpoint string on OpenRouter. Open-weight models set a floor: DeepSeek prices V4.1 Flash 1017 at 0.15 / 0.60 / 0.003, and the number of inference providers hosting it on OpenRouter has grown from 13 for V4 Flash in May to 29; Kimi K3.1 Pro is priced at 2 / 10 / 0.20, and once its weights land on November 4 more providers can host it. As long as third-party providers can offer similar quality at similar prices, the big labs will find it hard to raise prices.

He also mentions two other factors: inference infrastructure really has become more efficient, and both companies are rumored to be close to going public, which makes cutting prices now a puzzle; he wants to see the inference margins in their prospectuses. He adds that subscription allowances may be squeezed before API rates: when OpenAI released GPT-6.1 Sol, it halved the usage allowance of the $200 Codex tier. All of this is Woolf's personal analysis, the parts about IPOs and margins are speculation, and each provider's official pricing page remains the reference for prices.

via: Max Woolf's Blog: OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?