Prompt caching pricing explained: cached input tokens cost 75–98% less — here is the field, priced.
Prompt caching pricing is the quiet line in every API price list: when your prompt reuses the same context, the repeated tokens are billed at a cached input rate instead of the full input rate. Across the models we track, that discount runs from 75% to 98% — and the standard rate is so common it is almost a convention. This post prices the whole field from our own database.
By AI Pro League · Published
Prompt caching pricing rewards repetition. Agents, chatbots and document tools resend the same system prompt, instructions and reference text on every call; caching lets the provider keep that prefix ready and bill its tokens at the cached rate. Of the 59 models in our AI model pricing index, 37 publish a cached input rate, and every one of them discounts it — the median discount is 90%, and 30 of the 37 sit at 90% or deeper.
How prompt caching pricing works
Only input tokens can be cached, and only the repeated part of them: the shared prefix at the start of the prompt. Fresh input and all output are billed at the normal rates on our pricing pages. That is why the cached rate matters most to workloads with a large, stable context — a long system prompt, a codebase, a document set — and barely at all to one-off short prompts, where there is nothing repeated to cache. The exact eligibility rules are each provider’s own; the rates themselves are published on their pricing pages, for example DeepSeek’s API pricing, and we record the checked date on every row of our database.
The deepest cached-input discounts we track
| Model | Maker | Input $/1M | Cached $/1M | Discount | Price checked |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | DeepSeek | $0.30 | $0.006 | 98% | Oct 18, 2026 |
| DeepSeek V4.1 Flash 1017 | DeepSeek | $0.30 | $0.006 | 98% | Nov 1, 2026 |
| Claude Fable 5.1 | Anthropic | $10.00 | $0.25 | 98% | Oct 14, 2026 |
| DeepSeek V4 Pro (0813) | DeepSeek | $1.32 | $0.044 | 97% | Oct 14, 2026 |
| Claude Opus 5.5 | Anthropic | $4.00 | $0.20 | 95% | Oct 14, 2026 |
| Claude Sonnet 5.5 | Anthropic | $2.00 | $0.10 | 95% | Oct 14, 2026 |
The shape of the field is the story: discounts cluster hard at 90% — 30 of 37 models — which makes the outliers at both ends worth noticing. The shallowest discount we record is 75%, and the deepest is 98%.
Worked example: what a 98% cached discount is worth
Take the deepest discount in the table, DeepSeek V4.1 Flash: full input is $0.30 per million tokens, cached input is $0.006. Run a workload of 10 million input tokens where 8 million are the same repeated context and 2 million are fresh:
| Cost | |
|---|---|
| All 10M at the full input rate | $3.00 |
| 8M cached + 2M fresh | $0.65 |
| Saving | $2.35 |
The saving scales with the repeated share, not with the total bill: at a 90%-discount model the same workload saves less per token, and a workload with no repeated context saves nothing at all. Our API cost calculator prices the input/output mix; for the cached share, run the arithmetic above with the two rates from the model’s row in the table.
When prompt caching pricing does not help
- One-off prompts. Nothing repeats, so nothing is cached — the headline input rate is the rate you pay.
- Output-heavy work. Caching never touches output tokens, which are the expensive side of most price lists; a cheap cached rate cannot rescue a costly output rate. Compare both on the cheapest capable models table.
- Churning context. If the shared prefix changes every call, the cache is rebuilt instead of reused and the discount never lands.
For the monthly picture of headline rates, see AI price drops — October 2026 and the League table price gap; for capability at each price, the League standings rank these same models by score.