Prompt Caching, Fully Specified

Prompt caching is the richest cost mechanic OpenAI publishes, and the easiest to get wrong quietly: a mistake doesn't raise an error, it just produces a smaller discount than you expected.

How it bills

Every input token is billed at exactly one of three rates — uncached input, cached input, or cache write. A write is not a fee added on top. On GPT-5.6 and later, a written token costs 1.25x the uncached rate instead of that rate — or 2.5x if you ask for the 12-hour lifetime — and later reads cost 0.1x whichever lifetime wrote the entry, except on GPT-6.1 Sol and GPT-6.1 Terra, where reads cost 0.05x. On earlier models there is no write surcharge at all: a written token bills as ordinary input, which is cheap but never free.

Why the write price changes the math

Because writes cost more on GPT-5.6 and later, caching there pays off only once a prefix is read back. With the default lifetime, the published rates put that break-even below a single read on every model — 0.28 reads at the general read rate, a little lower on GPT-6.1 Sol and Terra. The 12-hour lifetime moves it out to 1.67 reads at the general rate: worth paying for a prefix you will read back over hours, not one a conversation reuses every minute anyway. OpenAI's own worked example checks the default case: at the general read rate, one write and one full reuse cost 1.35 times the prefix's ordinary input cost, against paying full input for both requests without caching. The per-model table above computes both break-evens from each model's own price cells.

What changed in how caching works

On GPT-5.6 and later, OpenAI routes cache traffic automatically: prompt_cache_key no longer improves hit rates and only separates cache accounting between your users. On earlier models a stable key still matters for routing. Lifetime changed shape too. Newer models keep an entry for at least 30 minutes after its last write or read by default, or — since October 20, 2026, in the Responses API — for at least 12 hours, chosen per request or per breakpoint with prompt_cache_options.ttl. Older models use in_memory or extended 24h retention instead. Two billing rules come with the longer lifetime: a token covered by several writes is billed once, at the longest lifetime among them, and upgrading an entry already cached for 30 minutes to 12 hours charges a fresh write with no credit for the first. GPT-5.6 and later also offer explicit breakpoints, a prewarm option that fills the cache without generating output, and Prompt Cache Diagnostics, which explains individual misses.

What invalidates a cache

A prefix that isn't byte-identical doesn't match. A timestamp, a request id, unstable key ordering, a per-user tool list, compaction of earlier turns — or on GPT-6, changing top-level reasoning effort mid-conversation instead of appending a configuration_update item — each makes requests unique from that point on.