OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?
B · Useful, limited A well-sourced pricing timeline and market argument by data scientist and blogger Max Woolf; the prices are checkable, but the explanation of why they fell is informed opinion, and margin figures come from unnamed reports.
Summary based on the full text·Original: about 12 minutes to read
In short
A blog post (25 October 2026) by Max Woolf, a data scientist who writes on AI tools. He lays out a June to September 2026 timeline of OpenAI and Anthropic price moves, input, output and cache-read prices per million tokens, and asks why two companies preparing for IPOs would cut their main revenue line by up to 80%. His answer: models are now easy to swap, open-weight models served by dozens of inference providers set a price floor, and some of the cut reflects cheaper serving. He argues the "prices will jump once the subsidy ends" view does not fit the evidence.
Key numbers
- 80%OpenAI's price cut on GPT-5.6 Luna on 30 July 2026, to $0.20 / $1.20 / $0.02 per million tokens
- 96% - 99%share of total token spend taken by cache reads when using agents for coding, per the author
- 43.3%cost per token on Vercel's AI gateway relative to April 2026, while aggregate spend rose
- 13 to 29distinct OpenRouter providers hosting DeepSeek V4 Flash in May against the V4.1 1017 variant in October
As reported by the source.
English·Web page·opens minimaxir.com
What it reports
- OpenAI cut GPT-5.6 Luna by 80% on 30 July; Anthropic made Sonnet 5's launch price of $2 / $10 / $0.20 permanent on 10 August.
- Fable 5.1 (1 September) kept Fable 5's $10 / $50 input and output price but cut cache reads from $1 to $0.25, a 75% cut. Because cache reads dominate agent coding spend, the author puts the effective cost cut at 70-75%.
- Claude Opus 5.5 (22 September) at $4 / $20 / $0.20 is 40-60% cheaper in total than Opus 5; Anthropic stated it "requires less compute to serve".
- GPT-6.1 Sol (29 September) at $2 / $10 / $0.10 makes that family cumulatively 75% cheaper.
- Open-weight pressure: DeepSeek V4.1 Flash 1017 scores 45 on the Artificial Analysis index against 24 for V4 Flash at a similar price, and Kimi K3.1 Pro scores 55 against Opus 5.5's 58 at half the input and output price.
- The author cut his own Codex subscription to a fifth after the price change; OpenAI cut the $200-a-month tier's usage allowance by 50% at the same time.
How they got it
Compiled from vendor announcements, OpenRouter provider listings retrieved on 24 October 2026, a Vercel gateway chart and the author's own subscription use. The economic argument is his own reasoning.
Limits
- The 96-99% cache-read share and 90-100% cache hit rates are stated as general figures, without the author's own measurement shown.
- Margin figures (profit-per-token "as healthy as 80%") come from unnamed reports with caveats.
- IPO timing is described as rumour.
- One observer's reading of a fast-moving market; prices may have changed since.
Why it matters for companies
Agent budgets set on mid-2026 prices are probably too high, and the line item that matters is cache reads, which vendor headlines often skip. The post supports two planning assumptions: frontier token prices are more likely to keep falling than to spike, and subscription allowances, not list prices, are where vendors will tighten first. Contracts should price cache reads explicitly and allow switching models without rework.
Summary based on the full text. Read the original for the full argument and data.Read the original on minimaxir.com