OpenAI and Anthropic lowered prices for several models between June and September 2026, according to a timeline by Max Woolf on Max Woolf’s Blog. Woolf’s clearest 80% example is OpenAI’s GPT-5.6 Luna. He lists its June 26 input, output and cache-read prices at $1, $6 and $0.10 per million tokens, respectively. On July 30, he lists them at $0.20, $1.20 and $0.02, an 80% reduction in each rate.
The size of the cut depends on the model
Woolf lists Anthropic’s Claude Opus 5 at $5 for input and $25 for output per million tokens in July, compared with $4 and $20 for Opus 5.5 in September. That is a 20% reduction in those rates, not an 80% cut. His account also describes an approximately 20% GPT-5.6 Sol discount announced in August as lasting three months.
Cache reads, which let a model reuse earlier material, have their own price. Woolf’s September comparison puts Anthropic’s Fable 5.1 cache-read rate at $0.25 per million tokens, down 75% from Fable 5’s $1 rate, while its input and output rates stayed at $10 and $50. Woolf estimates that cheaper cache reads could substantially lower coding-agent costs under his assumptions about how often those agents reuse earlier conversation.
Why Woolf thinks prices fell
Woolf suspects the Luna cut was intended to compete with smaller open-weight models. He also points to competition among companies that serve those models: in an OpenRouter comparison retrieved Oct. 24, 2026, he counted 29 providers hosting the DeepSeek V4.1 Flash 1017 variant. He argues that prices may fall further if providers can keep offering models of similar quality at similar prices.
Woolf considers a different possibility, too: improved infrastructure could help explain some reductions. He quotes Anthropic as saying Opus 5.5 requires less computing power to serve than Opus 5 and that its pricing reflects that. Anthropic’s explanation for Opus does not establish why the other prices in Woolf’s timeline changed.
Comments