Benchmarks

Prompt caching facts

How prompt caching behaves for each model on each platform: whether it caches, the smallest prompt that gets cached, how long the cache lasts, and what reads and writes cost.

Not promised means the provider gives no guarantee on how long the cache lasts.

exact
the provider promises this lifetime
at least
promised for at least this long
not promised
no guarantee from the provider

Amazon Bedrock

32 models

Azure

50 models

Google Vertex

40 models

Anthropic

19 models

OpenAI

42 models

Google AI Studio

11 models

xAI

9 models

Click a row for every number. Prices are USD per million tokens.

We measure these daily and do our best to keep them accurate. Always confirm with the provider’s official documentation.

Official docs: AnthropicOpenAIGoogle GeminiVertex AIAWS BedrockAzure

Spot a wrong number? Message our support.

Prompt caching facts · TryAii