OpenAI improves GPT-6 prompt caching with higher hit rates and explicit breakpoints

Alongside the GPT-6 Sol and Luna launch, OpenAI shipped a technical upgrade to GPT-6 prompt caching aimed squarely at developer economics. The changes include higher cache hit rates, explicit breakpoints that let developers mark where cacheable context ends, new diagnostics to see caching behavior, and finer controls that together reduce both latency and cost.

Prompt caching matters most for agentic and RAG-style workloads that resend large, stable context blocks (system prompts, tool schemas, documents) on every call. By raising hit rates and exposing explicit breakpoints, OpenAI lets developers guarantee more of that repeated context is served from cache rather than re-processed, which cuts per-call cost and speeds responses.

The strategic framing is consistent with the week's price war: on top of halving per-token prices for Sol and Luna, better caching further lowers the effective cost-per-task — the metric Sam Altman emphasized when he said the new models are 'even less per task.' For high-volume production features, caching gains can matter as much as headline token prices.

Competitively, prompt caching is now table stakes — Anthropic's Opus 5.5 and AWS's Strands harness both lean on caching and context management (Strands claims 28% lower token cost partly from these defaults). OpenAI's explicit-breakpoint control is a developer-experience differentiator, giving more determinism than implicit caching. The caveat: caching gains depend heavily on workload shape, so the real-world savings will vary, and developers will want to verify hit-rate claims with the new diagnostics. What to watch: whether the explicit breakpoint model becomes a de facto standard other providers copy.