OpenAI upgrades GPT-6 prompt caching with up to 90% discount on cached input

OpenAI announced improved prompt caching for GPT-6 on September 22, 2026. The update raises cache hit rates by default and adds a prompt caching dashboard, a diagnostics tool, explicit breakpoints, and prewarming. Under the new scheme, eligible shared prefixes that are reused within a 30-minute window can have their cached input tokens discounted by up to 90%. The dashboard tracks hit rates, while the diagnostics tool explains exactly why a given request missed the cache, with reasons such as tools_changed and token counts. At this point only a single public write-up is available, so the announcement stands alone and any follow-up details have not been reported. Overall, this looks like a straightforward cost-reduction and observability upgrade for teams that rely on repeated multi-turn exchanges: the default hit-rate improvement, explicit breakpoints, and prewarming make caching behavior easier to control, while the diagnostics tool should help surface what is causing misses. In practice, organizations using long conversation histories, large tool sets, or fixed prefixes may benefit the most, since those are the workloads where cache behavior tends to matter. The 90% ceiling applies only to eligible cached input tokens reused inside the 30-minute window, so results will depend on request patterns and on whether prefixes stay eligible. No performance or latency numbers were provided beyond the discount, and there is no independent verification of the claimed improvements yet.