OpenAI improves prompt caching for GPT-6

OpenAI has released a prompt caching update for the GPT-6 family: explicit cache stop points via the prompt_cache_options.mode and prompt_cache_breakpoint parameters, a Prompt Cache Diagnostics tool and dashboard, and a discount on cached input tokens of up to 90%.

Compiled by Sergey KostenchukPublished 2026-09-22Updated 2026-09-22

2026-09-22 Coding OpenAI

💻 OpenAI improves prompt caching for GPT-6

Cache hit rates have increased, explicit cache stop points (prompt_cache_options.mode and prompt_cache_breakpoint parameters), a Prompt Cache Diagnostics tool, and cache “warming” before a request are now available. The discount on cached input tokens reaches 90%: write — 1.25x the input price, read — 0.1x.

🌍 Caching is becoming a managed part of the API contract rather than a hidden server-side optimization: cache boundaries are set explicitly, and reasoning effort can be changed between responses without resetting the cache.

👤 Working with the OpenAI API? Check your hit rate on the Prompt Caching Dashboard and set explicit breakpoints. For repeated prompts of 1,024 tokens or more, this is a way to pay up to 10% of the regular price for input tokens and get the first response token faster.

Source 1: https://openai.com/index/better-prompt-caching-for-gpt-6 Source 2: https://developers.openai.com/api/docs/guides/prompt-caching

Sources