OpenAI adds cache breakpoints and prewarming to GPT‑6 prompt caching
The company said the changes raise hit rates and cut latency, and that reasoning changes no longer throw away a cached prefix.
OpenAI said it has changed how prompt caching works for GPT-6, adding higher cache hit rates, new diagnostics, explicit breakpoints and prewarming. The company published the changes on Tuesday evening, hours after launching GPT-6 Sol and Luna.
Rohan Paul, who summarises AI papers and releases, said the point is long-running agents that resend instructions, tool definitions and context on every turn. OpenAI's own page frames the changes the same way, as controls that reduce latency and cost.
The figures are not there. OpenAI's announcement names the features without putting a number on the hit-rate gain, the latency saving or the discount on cached tokens. Neither account gives a figure, and the paper has not tested the behaviour.
Prompt caching cuts what a developer pays to resend the same prefix, so the changes may matter most to agent builders running long tool loops. Whether the higher hit rates arrive in practice rests on OpenAI's description, which is the only account of them so far.