What a cache hit saves
Prompt caching prices
What a cache hit saves, and how many reuses pay for the write. Every figure is derived from the prices a host lists. Nothing here is measured. All prices.
1,153 models
149 of these models list a cache-read price and 128 list a cache-write price. Break-even is blank for the rest: without price_cache_write the write premium is unknown, and assuming it is zero would put a confident 0 in 21 rows nobody measured. There is no TTL column and no minimum-token column because no source we ingest publishes either field.
A worked example
One million prefix tokens on Claude Haiku 4.5 Thinking, sent once and reused 10 times. Cached costs $1.25 to write plus 10 reads at $0.10, so $2.25. Uncached costs 11 sends at $1.00, so $11.00. Break-even for this model is 1 later request.
Common questions
- What is prompt caching?
- A host stores the tokens at the start of your prompt. A later request that repeats that prefix is charged a cache-read rate instead of the input rate. The saving is on input tokens only. Output is charged as usual.
- How much does a cache hit save?
- The discount column is (input price − cache-read price) ÷ input price, from the prices each host lists. It is not a measurement. Most hosts list a cache read at one tenth of their input price.
- What is the break-even column?
- The smallest number of later requests at which caching a prefix costs less than sending it uncached every time. Writing once and reading n times costs (cache write + n × cache read). Not caching costs (1 + n) × input. The column is blank when the host lists no cache-write price, because the write premium is then unknown.
- Why is there no cache TTL or minimum-token column?
- No source this site ingests publishes either one. Checked on 2026-09-02 against every corpus row: the fields available are the cache-read price, the cache-write price, a longer-TTL write price at a few hosts, and a caching capability flag. A TTL column would be invented, so there is none.
- Do all hosts of one model charge the same cache price?
- No. A price belongs to a model and a host. This table shows one listing per model. Open the model page to see every host.