The lede
OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?
Max Woolf
Everyone building on these APIs has noticed the bills falling and quietly assumed it was loss-leader pricing with a reckoning attached; Woolf lines up the actual cuts — Luna down 80% in July, Sol halved in September, Opus 5.5 landing 40–60% under Opus 5 — and argues the reckoning theory does not survive contact with the substitutes. The mechanism worth internalizing is the cache read: it is 96–99% of agentic coding spend, which means Fable 5.1's 75% cache-read cut was effectively a 70–75% price cut on the only line item that matters, and it explains why the labs keep discounting the cheap-looking number instead of the headline one. Vercel's gateway data has cost per token at 43% of April's level while aggregate spend climbs anyway, which is the shape of a market where demand outruns deflation. He is honest about the part he cannot explain — the losses, the IPOs, the reported 80% per-token margins — and the useful conclusion is narrower than the headline: price increases look unlikely while OpenRouter makes switching a config change, but the subscription allowances are where the squeeze actually shows up first.