OpenAI, Anthropic cut LLM prices up to 80% on open-weight pressure

The price cuts did not happen in a vacuum. Woolf's analysis points to a structural feature of the LLM market: switching costs are low because tools like OpenCode and Zed Agent let developers swap models easily. That makes it hard for OpenAI or Anthropic to lock users in, and it shows up in the data — on Vercel's AI Gateway, 48.0% of teams changed their most-used model in September, and 28.6% changed the company that makes their main model.

Open-weight models reinforce that pressure. Because their weights are public, multiple providers can host the same model and compete on price. On OpenRouter, 29 providers offered DeepSeek V4.1 Flash 1017 as of October 24, 2026, some charging around $0.003 per million tokens for cache reads. DeepSeek's own cache-read price is $0.003 per million tokens, just 2% of its regular input price. Woolf sees OpenAI's 80% cut to GPT-5.6 Luna in July as a response to small open-weight models of the time.

Cheaper serving also played a role. Anthropic said Opus 5.5 needs less compute to serve than Opus 5, and OpenAI said efficiency gains across models, inference systems, and tool integration let it do more work with the same compute. Woolf argues that a strategy of gathering users and then raising prices sharply is hard to pull off, because users can move to a comparable model from another provider.