Prompt Caching

FastRouter supports prompt caching across all major providers — automatically where providers allow it, and with explicit controls for Anthropic Claude.

Overview

Prompt caching reduces the cost of repeated context — long system prompts, RAG chunks, documents — by charging a fraction of the normal input price on cache hits.

Zero-Config Providers

The following providers cache automatically. No changes to your requests needed.

Provider	Cache write	Cache read
OpenAI	Free	0.25x – 0.50x input
DeepSeek	Same as input	~0.10x input
Google AI Studio	Free	0.10x input
Google Vertex AI	Free	0.10x input
Grok	Free	See provider pricing
Moonshot AI	Free	See provider pricing
Baseten	Free	See provider pricing

OpenAI requires a minimum of 1024 tokens.

Google AI Studio and Vertex AI both support implicit caching on Gemini 2.5 and newer models — no configuration needed. FastRouter keeps your prompt prefixes stable to maximize cache hits. The 0.10x cache-read rate (90% discount) applies to all Gemini 2.5+ models; legacy Gemini 2.0 Flash is discounted at 0.25x. Implicit caches are managed entirely by Google's serving infrastructure with no storage cost to you. TTL is typically 3–5 minutes. To maximize cache hits, keep large static content (system instructions, RAG context, few-shot examples) at the beginning of your prompt and push dynamic content to the end.

Minimum token thresholds before caching applies:

Model	Min tokens
Gemini 2.5 Pro	4,096
Gemini 2.5 Flash	1,024
Gemini 2.5 Flash-Lite	1,024

---

Anthropic Claude

Anthropic requires you to explicitly mark what should be cached using cache_control. FastRouter supports two approaches.

Option A — Top-level (recommended for chat)

Add cache_control once at the request root. FastRouter automatically places the cache breakpoint at the last cacheable block and advances it as the conversation grows.

{
  "model": "anthropic/claude-sonnet-4.6",
  "cache_control": { "type": "ephemeral" },
  "messages": [...]
}

Only works when routed to Anthropic directly.

Option B — Per-block (for precise control)

Place cache_control on individual content blocks. Useful when you have a large stable payload (a document, RAG chunks, a character card) and want to cache exactly that. Maximum 4 breakpoints per request.

{
  "messages": [
    {
      "role": "system",
      "content": [
        { "type": "text", "text": "You are a research assistant." },
        {
          "type": "text",
          "text": "<large document>",
          "cache_control": { "type": "ephemeral" }
        }
      ]
    },
    { "role": "user", "content": "Summarize the findings." }
  ]
}

Per-block caching works across Anthropic and Vertex.

TTL

TTL	Syntax	Write cost	Read cost
5 min (default)	{ "type": "ephemeral" }	1.25x input	0.10x input

Model minimums

Min tokens	Models
4096	Opus 4.5, 4.6, 4.7 · Haiku 4.5
2048	Sonnet 4.6 · Haiku 3.5
1024	Sonnet 4, 4.5 · Opus 4, 4.1 · Sonnet 3.7

---

Checking Cache Savings

Every API response includes a prompt_tokens_details object:

"prompt_tokens_details": {
  "cached_tokens": 10318,
  "cache_write_tokens": 0
}

cached_tokens > 0 means you're hitting the cache.

You can also check per-request cache usage on the Activity Logs page flyout on the FastRouter dashboard.