Prompt caching

Prompt caching saves on a repeating request prefix (system instruction, history, tool descriptions). For supporting models a repeated prefix is read from the cache much cheaper than normal input — on the order of tens of percent of the full price. Cache support for each variant is shown on the Modelspage.

On /v1/chat/completions models of the OpenAI family cache automatically: an identical stable prefix on repeated requests is read from the cache. No special fields are needed — the main thing is not to change the start of the request. That is why the body is passed to the endpoint byte for byte, without reordering fields.

Claude works differently: it caches only when there are explicit cache_control breakpoints in the body — repeating the prefix does not switch caching on by itself. Put such a breakpoint at the end of the unchanged part (system or the last block):

curl https://getvibeapi.com/v1/messages \
  -H "x-api-key: $GETVIBE_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-5.5",
    "max_tokens": 256,
    "system": [
      {
        "type": "text",
        "text": "A large fixed instruction worth caching…",
        "cache_control": { "type": "ephemeral" }
      }
    ],
    "messages": [{"role": "user", "content": "Question about the instruction"}]
  }'

For Claude we support the native Anthropic format — /v1/messages with the x-api-key header. It is what Claude Code and the Anthropic SDK use: the body goes upstream as it is, with no translation from the OpenAI shape, so both the cache breakpoints and the cache_creation / cache_read fields in the response survive intact.

That is a measurement, not a promise: on 2026-08-21 an agent session on claude-opus-5 over /v1/messages — 119 requests, 15M input tokens — read 91.6% of its input from the cache. How much cache you actually get is always visible in the response usage and in the request statistics in your dashboard.