OpenAI improved prompt caching for GPT-6 with higher cache hit rates by default. It now discounts eligible shared prefixes reused within 30 minutes. New tools let developers monitor cache performance and diagnose misses. Explicit cache breakpoints allow choosing which prompt prefixes to reuse. Reasoning effort can now change mid-conversation without breaking cache. Tool definitions can be kept stable to preserve cache as instructions change. Prewarming prepares known context ahead of time to reduce latency. The Prompt Caching Dashboard tracks hit rates and input token composition. A diagnostics tool identifies changes that prevent reuse. GitHub Copilot reduced prompt tokens needing fresh processing by over 50%.
What changed
Cached prefixes were eligible for reuse only under stricter, shorter conditions.
What it unlocks
Reusing computation across requests to reduce response times and token costs for persistent agents.
- up to 90% discount on cached input tokens
- 30-minute cache eligibility window
- 50%+ reduction in prompt tokens requiring fresh processing for GitHub Copilot
- 85% to 90%+ cache hit rate improvement for Manus
What you need to act on it
- using GPT-6 models
- API requests within a 30-minute window
Sources