How much does prompt caching save on LLM API costs?
Published
Prompt caching cuts the price of input tokens a model has already processed, and it does not touch output tokens. As of Nov 22, 2026, 127 models in Sovyron's catalog list a cached-input rate on the maker's own API, with a median of 10% of the normal input price. The saving depends on the cache hit rate.
How much does a cache hit save?
Prompt caching lets a provider reuse the work it did on the start of an earlier prompt. A later request that begins with the same text pays a cached-input rate for those tokens. Only the new text after the shared start is billed at the normal input rate, and output tokens are billed as usual.
Across the 127 models with a cached-input price on the maker's own API, the median cache read costs 10% of the normal input price, and 64 of them sit at 10%. The lowest share is DeepSeek V4.1 Flash 1017 at 2%. The highest is GPT-4o mini at 50%. Anthropic's pricing page lists cache hits at 0.1x the base input price for most models and 0.05x for Claude Opus 5.5 and Claude Sonnet 5.5.
| Model | Input / 1M | Cache read / 1M | Read vs input | Cache write / 1M |
|---|---|---|---|---|
| GPT-6.2 Astra (OpenAI) | $8 | $0.4 | 5% | $10 |
| Gemini 4 Neon (Google) | $3 | $0.15 | 5% | none listed |
| GLM-5.3-FlashX (Z.AI) | $0.37 | $0.075 | 20.3% | none listed |
| Claude Fable 5.2 (Anthropic) | $8 | $0.2 | 2.5% | $10 |
| Mistral Large 4 (Mistral) | $0.7 | $0.07 | 10% | none listed |
| Qwen3.8 Omni Flash (Alibaba) | $0.15 | $0.016 | 10.7% | none listed |
| Muse Ember (Meta) | $0.1 | $0.004 | 4% | none listed |
Some providers also charge more to write a prompt into the cache. 40 of the 127 listings carry a cache-write rate, with a median of 1.25x the input price. OpenAI documents a 1.25x write rate for GPT-5.6 and later at the default 30-minute lifetime, and Anthropic documents 1.25x for a 5-minute write. A write that is never read back costs more than skipping the cache.
Worked example: a 20,000-token prompt sent 1,000 times
Take Claude Sonnet 5.5 on Anthropic. Input costs $2 per 1M tokens, a cache read $0.1 and a cache write $2.5. Say every request starts with the same 20,000-token system prompt and reference text, and 1,000 requests arrive. Without caching, those repeated tokens cost $40.00. In the table, every request that misses the cache pays the write rate.
| Cache hit rate | Without caching | With caching | Change |
|---|---|---|---|
| 0% | $40.00 | $50.00 | +25% |
| 20% | $40.00 | $40.40 | +1% |
| 50% | $40.00 | $26.00 | −35% |
| 80% | $40.00 | $11.60 | −71% |
| 95% | $40.00 | $4.40 | −89% |
The figures cover the repeated prefix only. New text and output tokens cost the same with or without caching. At these rates caching starts to pay once about 20.8% of requests hit the cache, which is (write rate - input rate) / (write rate - read rate).
When prompt caching does not pay off
- The start of the prompt changes. A hit needs the whole prefix to match. Anthropic caches tools, then system, then messages in that order, and OpenAI states that reuse requires the entire rendered prefix to match. A timestamp or a user name near the top breaks every match after it.
- Requests arrive too far apart. Anthropic gives a cache entry a 5-minute lifetime by default and refreshes it at no cost each time it is used. OpenAI documents a 30-minute default for GPT-5.6 and later. Traffic that comes less often pays the write rate again. Anthropic offers a 1-hour lifetime at a higher write price, and OpenAI offers 12 hours in the Responses API.
- The prompt is short. A prompt has to reach the model's minimum cacheable length before it can be cached. OpenAI documents 1,024 tokens for GPT-5.6 and later.
- Each prompt is used once. Where a write premium applies, a single-use prompt costs more with caching than without it.
How to get cache hits and check them
- Put everything that stays the same at the start of the prompt: tool definitions, the system prompt, reference documents and examples. Put the user's question and any changing data at the end.
- Keep the shared start identical from one request to the next. Serialize tool lists and JSON keys in the same order every time.
- Turn caching on where the provider asks you to. Anthropic opts in with cache_control, and OpenAI enables caching by default for supported models.
- Read the usage data in each response. OpenAI reports cached_tokens, and the usage object from Anthropic reports cache reads and writes. OpenAI also documents a Prompt Caching Dashboard and a Prompt Cache Diagnostics tool for finding misses.
- Divide cached requests by all requests to get your hit rate and compare it with the break-even above. The cost calculator on each model page, such as the Claude Sonnet 5.5 page, counts cached tokens at the cached-input price.
FAQ
Does prompt caching reduce the cost of output tokens?
No. Caching applies to input tokens the model has already processed. Output tokens are billed at the normal output rate, so the saving is largest on prompts that are long and repeated, with short answers.
Is prompt caching automatic, or do I have to turn it on?
It depends on the provider. OpenAI documents prompt caching as enabled by default for supported models. Anthropic documents it as opt-in through cache_control, either as automatic caching or with explicit breakpoints.
How long does a cached prompt stay cached?
Anthropic documents a 5-minute default lifetime that is refreshed at no cost each time the entry is used, with a 1-hour option at extra cost. OpenAI documents a 30-minute default for GPT-5.6 and later, with a 12-hour option in the Responses API.
Why is my prompt cache hit rate so low?
The usual cause is a prefix that differs between requests. Any change before the cached section, such as a timestamp, a user id or a reordered tool list, stops the match. Move changing content to the end of the prompt and check that requests come inside the cache lifetime.
Which LLM API has the cheapest cached input?
Among the makers' own APIs in the catalog, Muse Spark 1.2 Contributor has the lowest cache read at $0.002 per 1M tokens as of Nov 22, 2026. A low cache read does not make the whole bill low, so compare the full price on its model page and in the cheapest LLM API ranking.