Prompt Caching

Reduce repeated prompt cost when the selected model and upstream support cache controls.

Overview

Prompt caching lets a model reuse the processing of a large, unchanging prefix (a long system prompt, a document, a tool schema) across requests, so you pay less and get faster responses for the repeated part. Caching is model-specific: some providers expose explicit cache markers, some cache implicitly, and some don't support it at all.

Explicit cache control (Claude)

On Claude models, mark the reusable prefix with cache_control. Subsequent requests that share that prefix read it from cache at a reduced rate.

{
  "model": "anthropic/claude-sonnet-4.6",
  "max_tokens": 512,
  "system": [
    {
      "type": "text",
      "text": "<a long, reused style guide or document>",
      "cache_control": { "type": "ephemeral" }
    }
  ],
  "messages": [{ "role": "user", "content": "Summarize section 3." }]
}

Implicit caching (OpenAI / others)

Some models cache automatically — keep the stable part of your prompt at the very start so the unchanging prefix can be matched and reused. No special field is needed; you simply structure the prompt prefix-first.

Cache tokens in usage

When caching applies, the response usage reports cached tokens so you can verify it is working and track savings. Field names vary by provider (e.g. cached_tokens, cache_read_input_tokens).

JSON
{
  "usage": {
    "prompt_tokens": 2200,
    "completion_tokens": 120,
    "prompt_tokens_details": { "cached_tokens": 2048 }
  }
}

Model-dependent features

OneRouter keeps protocol compatibility and forwards model-specific parameters where supported. Features such as tool calling, structured output, reasoning, prompt caching, web search, image input, audio input, video generation, and provider-specific options depend on the selected model and upstream provider. Check the model detail page before relying on a feature in production.