OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why? · Max Woolf · personal blog, October 25, 2026
A practitioner's timeline shows agent token costs falling by up to 75% since June, driven by cache-read cuts
Machine-readable
Download Markdown
Woolf lists model launches and price changes from both labs between June 26 and September 29, 2026, quoting input, output and cache-read prices per million tokens. Examples he gives: Anthropic's Fable 5.1 cut cache-read cost by 75% against Fable 5 at the same input and output price; Opus 5.5 launched at $4 / $20 / $0.20, which he calls a 40 to 60% total reduction from Opus 5; and OpenAI's GPT-6.1 Sol, at $2 / $10 / $0.10, halved the cache price, which he describes as making that model family cumulatively 75% cheaper. His central claim is that cache reads make up 96 to 99% of token spend when agents code, with cache hit rates of 90 to 100%, so a 75% cache-read cut is roughly a 70 to 75% cut in what a session costs. That figure is his estimate; he does not show the measurement behind it.
He also cites Vercel's gateway data putting cost per token at 43.3% of its April 2026 level, while noting that aggregate spend keeps rising because models improve faster than they get cheaper. On open weights, he reports 29 providers now hosting DeepSeek V4.1 Flash on OpenRouter, up from 13 for the earlier version, and says Moonshot's Kimi K3.1 Pro scores 55 against Opus 5.5's 58 at half the input and output price. All prices are his compilation of vendor announcements; they were not checked against the vendors' own pages for this item.
His forecast is that prices keep falling, because switching models costs almost nothing and many providers can serve comparable models. He argues a sudden jump to a "true cost of a token" is unlikely. He also reports that OpenAI cut the Codex $200 tier's usage allowance by 50% when GPT-6.1 Sol launched, and expects subscription value to be cut before base prices if capital tightens. For his own use, he says he cut his Codex usage quota to one-fifth after the price reduction.
Why it matters: If agent cost is mostly cache reads, then cache-read price and your hit rate matter more than the headline input and output rates when you compare models for a harness. Check your own token breakdown before accepting that 96 to 99% figure, and be aware that the subscription tiers moved differently from API prices.