Prompt Caching

Prompt caching on Anthropic lets you cache individual messages in your request for repeat use. With caching, you can free up your tokens to include more context in your prompt, and also deliver responses significantly faster and cheaper.

You can use this feature on our OpenAI-compliant universal API.

​

API Support

Just set the cache_control param in your respective message body:

REST API

curl https://aigw.portkey.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $ANTHROPIC_API_KEY" \
  -H "x-portkey-api-key: $PORTKEY_API_KEY" \
  -d '{
    "model": "@anthropic/claude-3-5-sonnet-20240620",
    "max_tokens": 1024,
    "messages": [
        { "role": "system", "content": [
            {
                "type":"text","text":"You are a helpful assistant"
            },
            {
                "type":"text","text":"<TEXT_TO_CACHE>",
                "cache_control": {"type": "ephemeral"}
            }
        ]},
        { "role": "user", "content": "Summarize the above story for me in 20 words" }
    ]
  }'

​

Cache TTL Options

By default, the cache has a 5-minute lifetime that refreshes each time cached content is used. You can optionally specify a 1-hour TTL by adding the ttl field to cache_control:

{
  "cache_control": { "type": "ephemeral", "ttl": "1h" }
}

TTL	Write Cost	Best For
5m (default)	1.25x base input price	Prompts used more frequently than every 5 minutes
1h	2x base input price	Agentic workflows, long conversations where follow-ups may exceed 5 minutes

Cache reads cost 0.1x the base input token price regardless of TTL.

• The message you are caching needs to cross minimum length to enable this feature (1024 tokens for Claude 3.5 Sonnet and Claude 3 Opus, 2048 tokens for Claude 3 Haiku)
• You can mix both TTLs in the same request, but 1-hour entries must appear before 5-minute entries
• Up to 4 cache breakpoints per request

For more, refer to Anthropic’s prompt caching documentation here.

​

Automatic Caching

Instead of marking individual messages, you can have the gateway turn on Anthropic’s automatic caching for every request by setting auto_caching to true in your Config:

{
  "provider": "@anthropic-provider",
  "auto_caching": true
}

The gateway adds a top-level cache_control: { "type": "ephemeral" } to the request. A cache_control you set on the request yourself is kept as is. auto_caching is inherited by nested targets and can be overridden on a target. It applies to Claude models on Anthropic, Google Vertex AI, Azure AI Foundry, and Claude Platform on AWS for /chat/completions and /messages, including requests to the Responses API (/v1/responses) that are sent to Claude. It also applies to /messages on Amazon Bedrock (InvokeModel) and Bedrock Mantle (anthropic.* models only). Bedrock Converse is not covered.

​

Seeing Cache Results in Prisma AIRS AI Gateway

The AI Gateway automatically calculates the correct pricing for your prompt caching requests & responses based on Anthropic’s calculations here:

In the individual log for any request, you can also see the exact status of your request and verify if it was cached, or delivered from cache with two usage parameters:

• cache_creation_input_tokens: Number of tokens written to the cache when creating a new entry.
• cache_read_input_tokens: Number of tokens retrieved from the cache for this request.

Understanding Token Counts with Caching

The AI Gateway normalizes Anthropic’s response to the OpenAI format. In this format, prompt_tokens includes the cached tokens:

prompt_tokens = inputTokens + cache_read_input_tokens + cache_creation_input_tokens

This differs from Anthropic’s native format where inputTokens excludes cached tokens. The AI Gateway’s pricing calculation accounts for this by:

1. Subtracting cached tokens from prompt_tokens to get the base input token count
2. Applying the standard input token rate to base tokens
3. Applying the discounted cache read rate to cache_read_input_tokens
4. Applying the cache write rate to cache_creation_input_tokens

This ensures accurate cost calculation even though the token format is normalized.