OpenAI adds improved prompt caching to GPT-6, offering up to 90% input discounts
OpenAI announced an improved prompt caching system for GPT-6, enabling persistent agents to run for hours on tasks such as code refactoring and document creation. The system reuses shared context across API calls, reducing response times and giving developers discounts of up to 90% on cached input tokens.
Key points
- Developers receive up to 90% discounts on cached input tokens for GPT‑6 agents.
- Cache eligibility lasts 30 minutes; a new dashboard shows hit rates and token breakdowns.
- Prewarming, explicit breakpoints, and adjustable reasoning effort let developers keep cache while changing tools.
Cache eligibility now lasts a 30‑minute window for shared prefixes. A new Prompt Caching Dashboard displays hit rates, an input composition chart, and token breakdowns, while a diagnostics tool explains misses, reporting figures such as 5629 affected tokens in a sample miss. Developers can set explicit cache breakpoints, adjust reasoning effort without breaking cache, and keep tool definitions stable using allowedtools or toolchoice settings.
OpenAI also introduced prewarming, which lets applications load known instructions, tool schemas, or reference material before a user query, moving processing out of the latency window. The refreshed prompt caching guide and monitoring tools aim to help developers optimize cache hit rates and overall integration performance.
Model page: GPT-6 →
Better prompt caching for GPT-6
OpenAI · 22 September 2026
Loading the full article…
This text was published by OpenAI. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- OpenAI research model uses crypto wallet funds without permission to finish task · 2 src
- Gmail Live rolls out to paid Workspace users · 2 src
- Researchers introduce ARIA framework for correcting multi-agent AI errors in real time · 1 src
- OpenAI introduces Dots agent for asynchronous task delegation · 4 src
- PrismML releases Ternary Bonsai 2 models for local AI on phones and wearables · 1 src
Comments
via GitHub Discussions