Response Caching

When a request is byte-for-byte identical to one answered recently, AnyRouter returns the cached response with no upstream call and no charge. This differs from Prompt Caching, which still runs inference. Any change to messages, temperature or another field makes a request distinct.

It helps for deterministic requests (temperature: 0), embeddings, repeated identical prompts such as health checks and demos, and development loops around a fixed prompt.

Default

Deterministic chat/completions requests (temperature: 0 or unset), embeddings and speech requests are cached for about 5 minutes. Caching is never applied to zero-data-retention (ZDR) workspaces or to requests that opt out.

Control it

With the SDK:

import { AnyRouter } from "@anyr/sdk"

const client = new AnyRouter()

// Cache this exact response for 10 minutes.
await client.chat.completions.create(
  { model: "z-ai/glm-4.7-flash", messages: [{ role: "user", content: "What is the capital of France?" }] },
  { cacheTtl: 600 },
)

// Force a fresh upstream call.
await client.chat.completions.create(
  { model: "z-ai/glm-4.7-flash", messages: [{ role: "user", content: "Latest news?" }] },
  { skipCache: true },
)

With @anyr/ai-sdk-provider, pass the same options as providerOptions: { anyrouter: { cacheTtl: 600 } }.

With raw headers on any HTTP client:

cf-aig-cache-ttl accepts 60 seconds to 1 month (2,678,400 seconds); values outside are clamped.

curl https://anyrouter.dev/api/v1/chat/completions \
  -H "Authorization: Bearer $ANYROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -H "cf-aig-cache-ttl: 600" \
  -d '{"model": "z-ai/glm-4.7-flash", "messages": [{ "role": "user", "content": "What is the capital of France?" }]}'