On 22 September OpenAI set out a reworked prompt-caching system for its GPT‑6 models, aimed at agents that run for hours and resend the same instructions, tool definitions and history on every call. Shared prefixes now qualify for a cache discount when reused within a 30-minute window, and OpenAI says hit rates are higher by default.
Alongside it come a caching dashboard, a diagnostics tool that returns a machine-readable reason for a miss and an estimate of the tokens affected, explicit breakpoints that let developers choose where a cached prefix ends, the ability to raise or lower reasoning effort mid-conversation without invalidating the cache, and prewarming of known context before a user’s first question.
One number is worth pinning down. The caching post says discounts of up to 90% on cached input tokens; the Sol and Luna post published the same day states discounts of 90%. The qualifier is doing work, and only one of the two pages carries it.
The most concrete outside figure comes from GitHub’s chief product officer, Mario Rodriguez, who is quoted saying Copilot has cut the share of prompt tokens needing fresh processing by more than half over recent months, across billions of requests, relative to its previous baseline. For long-running agents, re-reading the same context is a large part of the bill.