OpenAI Updates Prompt Caching for GPT-6 With 90% Cheaper Cache Reads

On September 22, 2026, OpenAI updated its prompt caching alongside GPT-6 Sol and GPT-6 Luna, according to two reports published on DEV Community on November 2, 2026. The updates, relayed through a Thai-language piece and a German technical blog, discount reads of cached input tokens by 90%, raise the default cache hit rate, and stop two common kinds of changes that previously invalidated the cache. OpenAI also added a Prompt Caching Dashboard and diagnostic tooling; the German post mentions a cache dashboard and support for explicit breakpoints.

The reports quote OpenAI saying that its own researchers' median daily spend on coding agents exceeds $600, with the 90th percentile at $7,000 per day, and note that this covers internal lab tooling rather than production API traffic. One piece also cites GitHub running the change in production and seeing a more than 50% reduction in prompt tokens that had to be reprocessed, though the underlying measurement methodology is not described.

At this point the available information stops here: the figures and claims come from these two secondary write-ups, and the materials do not include OpenAI's own pricing page or a first-hand description of how cached reads are billed, so the exact per-token rate for cache reads remains unspecified in the sources reviewed.