As a conversation goes on, the AI model reads the earlier messages again, and providers charge less when it reads them from their cache. Anthropic charges a twentieth of the input price for these cached reads on Claude Sonnet 5.5. ToRun charged a tenth, so they now cost half as much as before. In GPT-5.5 and GPT-5.4 conversations longer than 272,000 tokens, cached reads were charged at the full input price. They are now charged at the cached price, a tenth of it.