Services or capabilities described in AWS documentation might vary by Region. To see the differences applicable to the AWS European Sovereign Cloud Region, see the AWS European Sovereign Cloud User Guide.

Prompt caching for faster model inference

Prompt caching is an optional feature that you can use with supported models on Amazon Bedrock to reduce inference response latency and input token costs. Amazon Bedrock supports two types of prompt caching: Implicit Prompt Caching and Explicit Prompt Caching. Support for each type varies by model and API.

Prompt caching can help when you have workloads with long and repeated contexts that are frequently reused for multiple queries. For example, if you have a chatbot where users can upload documents and ask questions about them, it can be time consuming for the model to process the document every time the user provides input. With prompt caching, you can cache the document so that future queries containing the document don't need to reprocess it.

Types of prompt caching

The two types differ in how reusable prompt content is selected:

Implicit Prompt Caching

Implicit Prompt Caching automatically attempts to reuse eligible prompt prefixes without requiring cache controls in your request. Keep static content at the beginning of your prompt and dynamic content at the end to increase the likelihood of an exact prefix match.

Implicit Prompt Caching is best effort. Repeating an identical prompt doesn't guarantee a cache hit, and cache-hit rates can vary.

Explicit Prompt Caching

Explicit Prompt Caching lets you opt in to caching with model-specific cache controls or cache checkpoints. Cache checkpoints are markers that define the contiguous subsection of your prompt that you want to cache. Prompt prefixes should remain static between requests. Changes to a prompt prefix in subsequent requests result in cache misses.

Cache checkpoints have a minimum and maximum number of tokens, depending on the model. You can only create a cache checkpoint if your total prompt prefix meets the minimum number of tokens. Minimums currently range from 512 to 4,096 tokens per cache checkpoint. For the minimum that applies to your model, see Supported models, Regions, and explicit caching limits or the model's model card. The minimum applies cumulatively to the entire prompt prefix before each checkpoint, including, where applicable, content in the tools, system, and messages fields. There is no minimum number of tokens required between cache checkpoints. For a model with a 1,024-token minimum, you can define additional checkpoints fewer than 1,024 tokens apart as long as the total prompt prefix before each checkpoint contains at least 1,024 tokens. If you add a cache checkpoint before the total prompt prefix meets the minimum number of tokens, your inference still succeeds, but your prefix isn't cached.

The cache has a Time To Live (TTL), which resets with each successful cache hit. During this period, the context in the cache is preserved. If no cache hits occur within the TTL window, your cache expires. Many models support a 5-minute TTL. Check the model card for your model to see the exact TTL conditions.

Explicit Prompt Caching provides control over which prompt content is eligible for caching. It doesn't guarantee that an eligible request results in a cache hit. A cache entry is stored on the server that processed the request, and a later request can read it only if that server handles it before the entry expires. Requests that are served by a different server, such as one in another Region, or that are sent before an identical request has started its response, write a new entry instead of reading one.

Billing for cached tokens

For both Implicit Prompt Caching and Explicit Prompt Caching, input tokens are billed at one of three rates. Tokens written to the cache are billed at the model's cache-write rate, which, depending on the model, can be higher than the standard input token rate. Tokens read from the cache are reported as cached tokens and billed at the model's cache-read rate. All other input tokens are billed at the standard input token rate. Implicit cache reuse on Amazon Nova models is an exception; see Supported models, Regions, and explicit caching limits. For more information, see the Amazon Bedrock pricing page.

You can use prompt caching when you run inference in Amazon Bedrock with supported models. Availability of each prompt caching type varies by model and API. Prompt caching is available through the following Amazon Bedrock features:

For models that support Explicit Prompt Caching, the APIs provide granular control over the prompt cache. You can set individual cache checkpoints within your prompts and add checkpoints up to the maximum allowed for the model. For more information, see Supported models, Regions, and explicit caching limits.

Supported models, Regions, and explicit caching limits

Prompt caching support varies by model and API. Model cards identify whether a model supports Implicit Prompt Caching, Explicit Prompt Caching, or both. Prompt caching is available in all AWS Regions where the supported models are available. To check model availability by Region, see Regional availability by models.

The following table lists models that support Explicit Prompt Caching, along with their token minimums, maximum number of cache checkpoints, and fields that allow cache checkpoints.

To see which prompt caching types a model supports, refer to Models at a glance, and then choose the model that you're interested in.

* Amazon Nova models support a maximum of 20K tokens for prompt caching. Prompt caching is primarily for text prompts.

To use the 1-hour TTL option, confirm in the preceding table that the model supports it. Check the model's model card for the latest model-specific TTL conditions. For supported models, specify the ttl field in your cache checkpoint. In the Converse API, add "ttl": "1h" to your cachePoint object. In the InvokeModel API for Claude models, add "ttl": "1h" to your cache_control object. For models that support only the 5-minute TTL, omit the ttl field. Including the field for these models can result in a ValidationException. If no ttl value is provided, the default 5-minute caching behavior applies. The 1-hour TTL is useful for longer-running sessions where you want to maintain the cache across extended periods.

Amazon Nova offers Implicit Prompt Caching for all text prompts, including User and System messages. This mechanism can provide latency benefits when prompts begin with repetitive parts, without explicit configuration. Implicit cache reuse on Amazon Nova models isn't reported in the cache usage fields and isn't billed as cache reads. Amazon Nova models that support Explicit Prompt Caching, listed in the preceding table, also let you specify cache checkpoints for more control over cache eligibility.

Prompt caching for models from Anthropic

Anthropic models that support prompt caching on Amazon Bedrock support Explicit Prompt Caching. Requests without cache controls aren't cached.

You can add cache controls in two ways:

Cache checkpoints

When a request includes a cache checkpoint, Claude also checks earlier content-block boundaries and reads the longest prefix that's already cached. The check finds only entries that earlier requests wrote, so a single checkpoint at the end of content that stays the same across requests, before any content that changes, is often enough.

For more control, you can use multiple cache checkpoints (up to 4 for Claude models) to set exact cache boundaries. Use multiple checkpoints when parts of your prompt change at different times. For example, place one checkpoint after your tool definitions, one after your system prompt, and one after a reference document. When only the document changes, the tool definitions and system prompt are still read from cache.

Automatic caching

Automatic caching is the simplest way to use Explicit Prompt Caching with Claude. You add one cache_control field at the top level of your request instead of placing cache checkpoints. Claude places the checkpoint on the last cacheable block and moves it forward as your conversation grows. Automatic caching uses one of the 4 checkpoint slots.

InvokeModel and Messages API request body:

{
  "anthropic_version": "bedrock-2023-05-31",
  "cache_control": {"type": "ephemeral"},
  "max_tokens": 1024,
  "system": "Long, unchanging system prompt...",
  "messages": [{"role": "user", "content": "First question"}]
}

Converse API request:

response = client.converse(
    modelId="us.anthropic.claude-sonnet-5-5",
    system=[{"text": "Long, unchanging system prompt..."}],
    messages=[{"role": "user", "content": [{"text": "First question"}]}],
    additionalModelRequestFields={"cache_control": {"type": "ephemeral"}}
)

Automatic caching is supported on Claude Haiku 4.5, Claude Sonnet 4.6, Claude Opus 4.7, Claude Opus 4.8, Claude Opus 5, Claude Opus 5.5, Claude Sonnet 5, Claude Sonnet 5.5, and Claude Haiku 5.5. For any other model, check the model card before you use automatic caching. Models that don't support it return a ValidationException for the top-level cache_control field; use cache checkpoints with those models.

Best practices for using cache management in Anthropic Models

If you have prompts that are used at a regular cadence (i.e., system prompts that are used more frequently than every 5 minutes), continue to use the 5-minute cache, since this will continue to be refreshed at no additional charge.

The 1-hour cache is best used in the following scenarios:

You can use both 1-hour and 5-minute cache controls in the same request, but with an important constraint: Cache entries with longer TTL must appear before shorter TTLs (i.e., a 1-hour cache entry must appear before any 5-minute cache entries).

Prompt caching for models from OpenAI

OpenAI models on Amazon Bedrock support Implicit Prompt Caching through supported APIs. GPT-5.6 and later models also support Explicit Prompt Caching. API availability varies by model and endpoint. Check the model's model card and the preceding table.

GPT-5.6 and later models

GPT-5.6 Sol (openai.gpt-5.6-sol), Terra (openai.gpt-5.6-terra), Luna (openai.gpt-5.6-luna), and later models such as GPT-6 Astra (openai.gpt-6-astra) and GPT-6.1 Sol (openai.gpt-6.1-sol) support both Implicit Prompt Caching and Explicit Prompt Caching. Explicit prompt cache breakpoints give you precise control over which portions of your prompt are eligible for caching. This is especially valuable for agentic workflows where system instructions, tool definitions, and reference files repeat across many calls while only the latest input changes.

Key characteristics:

Understanding the response

The usage object in the response includes two cache-specific fields:

When cached_tokens is greater than zero and cache_write_tokens is zero, your request fully matched an existing cache entry — no new writes occurred, and you received the maximum cost savings.

Best practices for using cache management in GPT-5.6 and later models

GPT-5.5 and earlier models

For OpenAI models prior to GPT-5.6 (such as openai.gpt-5.5 and openai.gpt-5.4), Implicit Prompt Caching is automatic. You don't need to add any special parameters. The system automatically attempts to cache eligible prompt prefixes of 1,024 tokens or longer. Cache writes have no additional fee on these models.

Key characteristics:

Best practices for using cache management in GPT-5.5 and earlier models

Getting started

The following sections show you a brief overview of how to use the prompt caching feature for each method of interacting with models through Amazon Bedrock.