Authors
Find out what your AI bill could drop to
Most of an AI coding bill is not answers. It is the same context sent again and again. Prompt caching is the provider's own discount for that, and most teams are only half using it.
What the cache is
Every request to a model carries a prompt: the instructions, the files it has read, the conversation so far. A coding agent sends most of that again on every turn. Prompt caching lets the provider remember the start of a prompt it has already processed, so the next request that begins the same way only pays full price for the new part.
Anthropic's documentation describes two ways to use it: automatic caching, where one setting on the request tells the system to cache the growing history and move the breakpoint forward as the conversation grows, and explicit breakpoints, placed on specific blocks for fine control (Anthropic, prompt caching documentation, accessed 19 October 2026).
What it costs, and what it saves
The numbers are on the providers' price lists. On Claude Opus 5.5 and Sonnet 5.5, writing to the cache costs 25 percent more than a normal input token, and reading from it costs 5 percent of the normal input price. On OpenAI models, cached input is billed at 10 percent of the input price. So a prompt that costs $1 to send fresh costs about $0.05 to send again from a warm Claude cache. For an agent that resends 100,000 tokens of context on every turn, that is the difference between paying for 100,000 tokens and paying for about 5,000.
Why it goes cold
The discount only applies while the cache is warm and the prompt still begins the same way. Three things break that, and all three are common.
- Time. The default cache lifetime is five minutes, refreshed on each use. A one-hour option exists at a higher write price. A developer who reads a ticket for ten minutes and comes back pays full price to rebuild everything.
- Anything changing early in the prompt. The cache matches from the start. A tool that reorders the system prompt, injects the date, or rotates a header invalidates every token after that point on every request.
- Compaction. When a long session is summarised to fit the context window, the prompt changes shape and the cache rebuilds. Claude Code's own usage screen now reports cache misses, expected rebuilds and whether the cache is warm, which tells you how often this happens to your team.
What a warm cache is worth
Take a team whose monthly bill is mostly context: say 60 percent of input tokens could be cached and today only half of them are. Moving the other half from full price to 5 percent takes roughly 28 percent off the input side of the bill, with no change to what anyone does and no change to the answers. That is one lever of several, and it is the one with the least to argue about, because the provider is the one offering the discount.
A proxy that sits in front of every tool can keep the cache warm on purpose: it orders prompts so the stable part comes first, keeps the breakpoints in the same place from one tool to the next, and chooses the one-hour lifetime where a team's rhythm needs it. It also sees the misses across the whole company, not one session at a time. If you want to know how warm your own cache is, that is the first thing we measure in a free bill review.
