Prompt Caching on Bedrock

Prompt caching on Amazon Bedrock lets you cache specific portions of your requests for repeated use. This feature significantly reduces inference response latency and input token costs by allowing the model to skip recomputation of previously processed content.

With Portkey, you can easily implement Amazon Bedrock’s prompt caching through our OpenAI-compliant unified API and prompt templates.

Model Support

Amazon Bedrock prompt caching is generally available with the following models:

Currently Supported Models:

• Claude 3.7 Sonnet
• Claude 3.5 Haiku
• Amazon Nova Micro
• Amazon Nova Lite
• Amazon Nova Pro

Customers who were given access to Claude 3.5 Sonnet v2 during the prompt caching preview will retain their access, but no additional customers will be granted access to prompt caching on the Claude 3.5 Sonnet v2 model.

How Bedrock Prompt Caching Works

When using prompt caching, you define cache checkpoints - markers that indicate parts of your prompt to cache. These cached sections must be static between requests; any alterations will result in a cache miss.

You can also use Bedrock Prompt Caching Feature with Portkey’s Prompt Templates.

Implementation Examples

Here’s how to implement prompt caching with Portkey:

NodeJS

import Portkey from 'portkey-ai'

const portkey = new Portkey({
    apiKey: "PORTKEY_API_KEY", // defaults to process.env["PORTKEY_API_KEY"]
    provider:"@PROVIDER" // Your Bedrock Virtual Key
})

const chatCompletion = await portkey.chat.completions.create({
    messages: [
        { "role": 'system', "content": [
            {
                "type":"text","text":"You are a helpful assistant"
            },
            {
                "type":"text","text":"This is a large document I want to cache...",
                "cache_control": {"type": "ephemeral"}
            }
        ]},
        { "role": 'user', "content": 'Summarize the above document for me in 20 words' }
    ],
    model: 'anthropic.claude-3-7-sonnet-20250219-v1:0'
});

console.log(chatCompletion.choices[0].message.content);

Supported Features and Limitations

Supported Features

• Text prompts and images embedded within text prompts
• Multiple cache checkpoints per request
• Caching in system prompts, messages, and tools fields (model-dependent)

Supported Models and Limits

Below is a detailed table of supported models, their minimum token requirements, maximum cache checkpoints, and fields that support caching:

Model	Model ID	Min tokens per checkpoint	Max checkpoints per request	Cacheable fields
Claude 3.7 Sonnet	anthropic.claude-3-7-sonnet-20250219-v1:0	1,024	4	system, messages, tools
Claude 3.5 Haiku	anthropic.claude-3-5-haiku-20241022-v1:0	2,048	4	system, messages, tools
Amazon Nova Micro	amazon.nova-micro-v1:0	1,000	4	system, messages
Amazon Nova Lite	amazon.nova-lite-v1:0	1,000	4	system, messages
Amazon Nova Pro	amazon.nova-pro-v1:0	1,000	4	system, messages

• The Amazon Nova models support a maximum of 32k tokens for prompt caching.
• For Claude models, tools caching is fully supported.
• Tools caching is not supported for Amazon Nova models.

Understanding Token Counts and Pricing

Portkey automatically calculates correct pricing for prompt caching requests. In the logs, you’ll see cache-related token counts in the usage object:

• 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.

Token Format Normalization

Portkey normalizes responses 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 native provider formats where input tokens may exclude cached tokens. Portkey’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.

Related Resources

AWS Bedrock Prompt Caching Docs

For more detailed information on Bedrock prompt caching, refer to: