OpenAI upgrades GPT-6 prompt caching: 0.1x read pricing, up to 78.5% savings on repeated inputs
新智元 · wechat · 2026-09-24
On Sep 22, OpenAI announced a major prompt caching upgrade for GPT-6, raising default hit rates and adding monitoring, diagnostics, and control tools.
- Pricing: for GPT-5.6+ models, cache writes cost 1.25x normal input; cache reads cost 0.1x. Reusing one document 10 times with 9 hits saves 78.5% on that repeated input. Caches require exact shared-prefix matches (≥1024 tokens for GPT-5.6+) and persist at least 30 minutes.
- How to keep hits: put static content first and append variable content; keep tool definitions and ordering stable; adjust reasoning effort incrementally rather than globally.
- New tooling: a cache dashboard breaking down hit rates; a diagnostics tool that pinpoints invalidation causes (one renamed tool cost 5,629 reusable tokens); and warmup, which pre-writes system prompts to cut time-to-first-token.
- Reported results: Manus's hit rate rose from 85% to 90%+; GitHub Copilot cut re-processed prompt tokens by over 50%.
More from Infra
- Omarchy 4.6 lets you share local AI models across your tailnet, says DHH — alexcovo_eth · 2026-11-12
- Eric Schmidt's foundation funds French NGO fighting against data centers — StephanSturges · 2026-11-12
- Epoch: Open-Weight GLM 5.3 Hacks Six Firefox Versions for ~$1,000 Each; Claude Fable 5.2 Sets ECI Record — Epoch AI · 2026-11-12
- Polymarket prices 28% odds of a state data center moratorium; Torus raises $500M — Polymarket · 2026-11-12
- AWS-wrapper cloud startup shuts down just one month after seed round — thedealdirector · 2026-11-12
- Utah startup Torus raises $500M for shipping container-sized AI data centers deployable in 4 months — Polymarket · 2026-11-12