Prompt caching: who offers it, and what it actually saves

Most of a real workload's input tokens repeat — a system prompt, a retrieved document, a few-shot preamble. Providers that cache those tokens bill them at a fraction of the normal input price on the next call. Across this index,364 of 539 priced endpoints publish a cached-input price, spread over 39 providers. Every figure below is computed from those published prices — nothing here is hand-written.

A worked example

Take a RAG / long context job — 8,000 input /500 output tokens per request, 1,000 requests — on Claude Fable 5.1 (Anthropic) (Anthropic), where 80% of the input is reused context and hits the cache:

Full price: $105.00  →  with caching: $42.60.
−$62.40 (59% cheaper)

Cache hit rate:80%  →  cost$42.60(59% saved)

The higher your reuse, the more caching pays off. Below ~30% hit rate the gain is marginal; a RAG or agent loop that re-sends the same context every turn is where it compounds.

Biggest input savings from caching

Endpoints ranked by how much their cached-input price undercuts their normal input price.

ModelProviderInput /1MCached /1MSaved
Claude Fable 5.1 (Anthropic)Anthropic$10.000$0.250−98%
Claude Fable 5.1 (Amazon Bedrock)Amazon Bedrock$10.000$0.250−98%
Claude Fable 5.1 (Google Vertex)Google$10.000$0.250−98%
DeepSeek V4 Flash 0731 (Fireworks)Fireworks AI$0.220$0.007−97%
DeepSeek V4 Flash 0731 (DeepSeek)DeepSeek$0.220$0.007−97%
DeepSeek V4 Flash 0731 (GMICloud)GMICloud$0.352$0.011−97%
DeepSeek V4 Flash 0731 (Cloudflare)Cloudflare$0.440$0.014−97%
DeepSeek V4 Flash Vision Exp (Fireworks)Fireworks AI$0.220$0.007−97%
DeepSeek V4 Flash Vision Exp (DeepSeek)DeepSeek$0.220$0.007−97%
DeepSeek V4 Pro 0813 (GMICloud)GMICloud$1.122$0.037−97%
DeepSeek V4 Pro 0813 (Parasail)Parasail$1.320$0.044−97%
DeepSeek V4 Pro 0813 (DigitalOcean)DigitalOcean$1.320$0.044−97%
DeepSeek V4 Pro 0813 (SiliconFlow)SiliconFlow$1.320$0.044−97%
DeepSeek V4 Pro 0813 (Cloudflare)Cloudflare$1.320$0.044−97%
DeepSeek V4 Pro 0813 (Fireworks)Fireworks AI$1.320$0.044−97%
DeepSeek V4 Pro 0813 (CoreWeave)CoreWeave$1.310$0.044−97%
DeepSeek V4 Flash 0731 (Novita)Novita AI$0.409$0.026−94%
DeepSeek V4 Flash 0731 (Reka)Reka$0.110$0.007−94%
DeepSeek V4 Flash 0731 (Phala)Phala$0.440$0.028−94%
DeepSeek V4 Flash 0731 (AtlasCloud)AtlasCloud$0.440$0.028−94%

Savings by provider

Median input discount and how many cached endpoints each provider publishes.

ProviderMedian input discountCached endpoints
DeepSeek−97%3
Mistral AI−90%10
Fireworks AI−90%9
Google−90%7
Amazon Bedrock−90%7
Chutes−90%6
Azure OpenAI−90%5
Anthropic−90%4
Minimax−90%4
SambaNova−90%1
xAI−85%1
Modal−84%4
BaseTen−83%10
Moonshot AI−83%3
Z.AI−82%10
Cloudflare−82%8
OpenAI−82%6
Reka−82%5
Morph−82%4
Novita AI−81%27
GMICloud−81%14
SiliconFlow−81%13
Alibaba−81%9
Together AI−81%7
Baidu−81%7
Friendli−81%6
DeepInfra−80%29
Venice−80%23
DigitalOcean−80%15
Relace−78%2
Phala−73%14
AkashML−63%6
Parasail−62%25
AtlasCloud−59%21
Io Net−53%5
Crusoe−50%7
Groq−50%5
CoreWeave−23%20
Cerebras−0%2

Prices per 1M tokens (USD), from each provider's published cached-input rate. Verified 2026-09-05.

FAQ

Which LLM API providers offer prompt caching?

In this index, 364 of 539 priced endpoints publish a cached-input price, spread across 39 providers. Coverage and the size of the discount vary by provider.

How much does prompt caching actually save?

On a RAG / long context workload on Claude Fable 5.1 (Anthropic) (Anthropic) where 80% of the input is reused context, caching cuts the bill by 59%. Savings scale with your cache-hit rate — below roughly 30% reuse the gain is marginal.

What kind of workload benefits most from caching?

Anything that re-sends the same context on every call — RAG over a fixed corpus, an agent loop replaying its history, or a long system prompt or few-shot preamble. One-off, unique prompts see little benefit because there is nothing to reuse.