Monitor your token usage by counting tokens before running inference

When you run model inference, the number of tokens that you send in the input contributes to the cost of the request and towards the quota of tokens that you can use per minute and per day. The CountTokens API helps you estimate token usage before sending requests to foundation models by returning the token count that would be used if the same input were sent to the model in an inference request.

Token counting is model-specific because different models use different tokenization strategies. The token count returned by this operation will match the token count that would be charged if the same input were sent to the model to run inference.

You can use the CountTokens API to do the following:

Supported models and Regions for token counting

To see which models support token counting, please visit models at a glance and pick the model you are interested in.

Count tokens using the bedrock-runtime endpoint

To count the number of input tokens in an inference request, send a CountTokens request with an Amazon Bedrock runtime endpoint, Specify the model in the header and the input to count tokens for in the body field. The value of the body field depends on whether you're counting input tokens for an InvokeModel or Converse request:

Example: count tokens for a bedrock-runtime request

The examples in this section let you count tokens for an InvokeModel and Converse request with Anthropic Claude 3 Haiku.

To try out counting tokens for an InvokeModel request, run the following Python code:

import boto3
import json

bedrock_runtime = boto3.client("bedrock-runtime")

input_to_count = json.dumps({
    "anthropic_version": "bedrock-2023-05-31",
    "max_tokens": 500,
    "messages": [
        {
            "role": "user",
            "content": "What is the capital of France?"
        }
    ]
})

response = bedrock_runtime.count_tokens(
    modelId="anthropic.claude-3-5-haiku-20241022-v1:0",
    input={
        "invokeModel": {
            "body": input_to_count
        }
    }
)

print(response["inputTokens"])

To try out counting tokens for a Converse request, run the following Python code:

import boto3
import json 

bedrock_runtime = boto3.client("bedrock-runtime")

input_to_count = {
    "messages": [
        {
            "role": "user",
            "content": [
                {
                    "text": "What is the capital of France?"
                }
            ]
        },
        {
            "role": "assistant",
            "content": [
                {
                    "text": "The capital of France is Paris."
                }
            ]
        },
        {
            "role": "user",
            "content": [
                {
                    "text": "What is its population?"
                }
            ]
        }
    ],
    "system": [
        {
            "text": "You're an expert in geography."
        }
    ]
}

response = bedrock_runtime.count_tokens(
    modelId="anthropic.claude-3-5-haiku-20241022-v1:0",
    input={
        "converse": input_to_count
    }
)

print(response["inputTokens"])

Count tokens using the bedrock-mantle endpoint

The bedrock-mantle endpoint exposes Anthropic's count_tokens API at /anthropic/v1/messages/count_tokens. Use it to count input tokens for Anthropic Claude models that don't support CountTokens on bedrock-runtime – for example, when the model is offered only through cross-Region inference (CRIS) on bedrock-runtime and so has no Region-specific endpoint for CountTokens to target. The /anthropic/v1/messages path is Claude-specific; non-Anthropic models on bedrock-mantle return The model 'X' does not support the '/anthropic/v1/messages' API.

The following example uses curl with a Amazon Bedrock API key to count tokens on the bedrock-mantle endpoint:

curl -X POST https://bedrock-mantle.us-east-1.api.aws/anthropic/v1/messages/count_tokens \
  -H "x-api-key: $BEDROCK_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic.claude-opus-4-7",
    "messages": [
        {"role": "user", "content": "How many tokens is this prompt?"}
    ]
  }'

The response contains an input_tokens field whose value is the token count for the supplied input.