Prompt Caching
What is prompt caching?
Prompt caching is a provider feature that lets a model reuse the work it already did on a stable prefix of your prompt, instead of reprocessing (and fully re-billing) that text on every call. You mark or structure the unchanging part — a system prompt, project conventions, reference material — so that repeated requests hit the cache and only the new tokens are charged at the full rate.
It doesn't make a single request smaller; it makes repeated requests that share a common prefix much cheaper. The savings compound over a session.
Why prompt caching matters in 2026
It matters because agentic coding sessions are long and repetitive: the same system prompt and project rules are sent on turn after turn. Without caching, that fixed context is re-billed every step. With it, the stable prefix is billed at a small fraction of the input rate after the first call, which is why keeping your instructions stable and cache-friendly is a real token lever — see Reduce Claude Code token usage.
The numbers, as of October 2026 (Anthropic and OpenAI pricing pages):
- Anthropic: a 5-minute cache write costs 1.25× the base input price, a 1-hour write 2×. A cache hit costs 0.1× on most Claude models, 0.05× on Opus 5.5 ($0.20/MTok against $4 input) and 0.025× on Fable 5.1.
- OpenAI: cached input on GPT-6 Astra is $1 against $10 for fresh input, i.e. 0.1×.
- Rate limits too: Anthropic's docs state that tokens read from cache do not count toward the input-tokens-per-minute limit on most models, so a high hit rate also buys rate-limit headroom.
Note what caching does not do: the cached tokens still occupy the context window and are still billed on every turn, just at the discounted rate. Caching makes a big prefix cheaper; it doesn't make it smaller.
When prompt caching doesn't help
- When the prefix keeps changing — if you rewrite your system prompt or reorder context each turn, there's nothing stable to cache.
- For one-shot prompts — a single request with no repetition has no prefix to reuse.
- When hits are rare — a write costs more than plain input (1.25× or 2× on Claude), so a prefix written and then never read again within its lifetime costs more than not caching at all.