MensajesGestión del contexto

Conteo de tokens

Cuenta los tokens de un mensaje antes de enviarlo a Claude. Usa los conteos de tokens para gestionar los límites de velocidad y los costos, tomar decisiones de enrutamiento de modelos y ajustar los prompts a una longitud objetivo.

El “token counting” (conteo de tokens) te permite determinar la cantidad de tokens en un mensaje antes de enviarlo a Claude. Esto te ayuda a tomar decisiones informadas sobre tus prompts y tu uso. Con el conteo de tokens, puedes:

Gestionar de forma proactiva los “rate limits” (límites de velocidad) y los costos
Tomar decisiones inteligentes de enrutamiento de modelos
Optimizar los prompts para una longitud específica

Cómo contar los tokens de un mensaje

El endpoint de conteo de tokensVista previa disponible con enlaces a la referencia de la API — presiona Tab hasta que el foco entre a la tarjeta y Escape para volver. acepta la misma lista estructurada de entradas que se usa para crear un mensaje, incluida la compatibilidad con “system prompts” (indicaciones del sistema), herramientas, imágenes y PDFs. La respuesta contiene el número total de tokens de entradaVista previa disponible con enlaces a la referencia de la API — presiona Tab hasta que el foco entre a la tarjeta y Escape para volver..

Count tokens in a Message

POST/v1/messages/count_tokens

Count the number of tokens in a Message.

messages: array of MessageParam

Input messages.

Our models are trained to operate on alternating user and assistant conversational turns. When creating a new Message, you specify the prior conversational turns with the messages parameter, and the model then generates the next Message in the conversation. Consecutive user or assistant turns in your request will be combined into a single turn.

Each input message must be an object with a role and content. You can specify a single user-role message, or you can include multiple user and assistant messages.

If the final message uses the assistant role, the response content will continue immediately from the content in that message. This can be used to constrain part of the model's response. This is called prefill. On models that don't support prefill, creating a message that ends with a partial assistant response returns a 400 error. See Prefill not supported.

Example with a single user message:

[{"role": "user", "content": "Hello, Claude"}]

Example with multiple conversational turns:

[
  {"role": "user", "content": "Hello there."},
  {"role": "assistant", "content": "Hi, I'm Claude. How can I help you?"},
  {"role": "user", "content": "Can you explain LLMs in plain English?"},
]

Example with a partially-filled response from Claude, for models that support prefill:

[
  {"role": "user", "content": "What's the Greek name for Sun? (A) Sol (B) Helios (C) Sun"},
  {"role": "assistant", "content": "The best answer is ("},
]

Each input message content may be either a single string or an array of content blocks, where each block has a specific type. Using a string for content is shorthand for an array of one content block of type "text". The following input messages are equivalent:

{"role": "user", "content": "Hello, Claude"}

{"role": "user", "content": [{"type": "text", "text": "Hello, Claude"}]}

See input examples.

Note that if you want to include a system prompt, you can use the top-level system parameter.

There is a limit of 100,000 messages in a single request.

One of the following:

UserMessageParam object{ content, role }

content: MessageContentParam

One of the following:

role: "user"

AssistantMessageParam object{ content, role }

content: MessageContentParam

One of the following:

role: "assistant"

SystemMessageParam object{ content, role }

model: Model

The model that will complete your prompt.

See models for additional details and options.

One of the following:

cache_control: optional CacheControlEphemeral { type: "ephemeral", ttl } or null

Top-level cache control automatically applies a cache_control marker to the last cacheable block in the request.

type: "ephemeral"

ttl: optional "5m" or "1h"

The time-to-live for the cache control breakpoint.

This may be one the following values:

5m: 5 minutes
1h: 1 hour

Defaults to 5m. See prompt caching pricing for details.

One of the following:

"5m"

"1h"

output_config: optional OutputConfig { effort, format }

Configuration options for the model's output, such as the output format.

effort: optional "low" or "medium" or "high" or 2 more or null

How much effort the model should put into its response. Higher effort levels may result in more thorough analysis but take longer.

Valid values are low, medium, high, xhigh, or max.

One of the following:

"low"

"medium"

"high"

"xhigh"

"max"

format: optional JSONOutputFormat { type: "json_schema", schema } or null

A schema to specify Claude's output format in responses. See structured outputs

type: "json_schema"

schema: map[unknown]

The JSON schema of the format

speed: optional "standard" or "fast" or null

The inference speed mode for this request. "fast" enables high output-tokens-per-second inference.

Some models require the fast-mode-2026-02-01 beta header for any speed value. See fast mode for which models.

One of the following:

"standard"

"fast"

system: optional string or array of TextBlockParam

System prompt.

A system prompt is a way of providing context and instructions to Claude, such as specifying a particular goal or role. See our guide to system prompts.

One of the following:

string

array of TextBlockParam { type: "text", text, cache_control, citations }

type: "text"

text: string

minLength1

cache_control: optional CacheControlEphemeral { type: "ephemeral", ttl } or null

Create a cache control breakpoint at this content block.

type: "ephemeral"

ttl: optional "5m" or "1h"

The time-to-live for the cache control breakpoint.

This may be one the following values:

5m: 5 minutes
1h: 1 hour

Defaults to 5m. See prompt caching pricing for details.

One of the following:

"5m"

"1h"

citations: optional array of TextCitationParam or null

One of the following:

CitationCharLocationParam object{ type: "char_location", cited_text, document_index, 3 more }

type: "char_location"

cited_text: string

document_index: number

minimum0

document_title: string or null

minLength1

maxLength500

end_char_index: number

start_char_index: number

minimum0

CitationPageLocationParam object{ type: "page_location", cited_text, document_index, 3 more }

type: "page_location"

cited_text: string

document_index: number

minimum0

document_title: string or null

minLength1

maxLength500

end_page_number: number

start_page_number: number

minimum1

CitationContentBlockLocationParam object{ type: "content_block_location", cited_text, document_index, 3 more }

type: "content_block_location"

cited_text: string

The full text of the cited block range, concatenated.

Always equals the contents of content[start_block_index:end_block_index] joined together. The text block is the minimal citable unit; this field is never a substring of a single block. Not counted toward output tokens, and not counted toward input tokens when sent back in subsequent turns.

document_index: number

minimum0

document_title: string or null

minLength1

maxLength500

end_block_index: number

Exclusive 0-based end index of the cited block range in the source's content array.

Always greater than start_block_index; a single-block citation has end_block_index = start_block_index + 1.

start_block_index: number

0-based index of the first cited block in the source's content array.

minimum0

CitationWebSearchResultLocationParam object{ type: "web_search_result_location", cited_text, encrypted_index, 2 more }

type: "web_search_result_location"

cited_text: string

encrypted_index: string

title: string or null

minLength1

maxLength512

url: string

minLength1

CitationSearchResultLocationParam object{ type: "search_result_location", cited_text, end_block_index, 4 more }

type: "search_result_location"

cited_text: string

The full text of the cited block range, concatenated.

end_block_index: number

Exclusive 0-based end index of the cited block range in the source's content array.

Always greater than start_block_index; a single-block citation has end_block_index = start_block_index + 1.

search_result_index: number

0-based index of the cited search result among all search_result content blocks in the request, in the order they appear across messages and tool results.

Counted separately from document_index; server-side web search results are not included in this count.

minimum0

source: string

start_block_index: number

0-based index of the first cited block in the source's content array.

minimum0

title: string or null

thinking: optional ThinkingConfigParam

Configuration for Claude's thinking.

With {"type": "adaptive"}, Claude decides when and how much to think. With {"type": "enabled"} (manual extended thinking), you set a budget_tokens of at least 1,024. Thinking tokens count toward your max_tokens limit.

Which type values are accepted, and what happens when you omit thinking, depend on the model. See thinking for each model's behavior.

One of the following:

tool_choice: optional ToolChoice

How the model should use the provided tools. The model can use a specific tool, any available tool, decide by itself, or not use tools at all.

One of the following:

tools: optional array of MessageCountTokensTool

Definitions of tools that the model may use.

If you include tools in your API request, the model may return tool_use content blocks that represent the model's use of those tools. You can then run those tools using the tool input generated by the model and then optionally return results back to the model using tool_result content blocks.

There are two types of tools: client tools and server tools. The behavior described below applies to client tools. For server tools, see their individual documentation as each has its own behavior (e.g., the web search tool).

Each tool definition includes:

name: Name of the tool.
description: Optional, but strongly-recommended description of the tool.
input_schema: JSON schema for the tool input shape that the model will produce in tool_use output content blocks.

For example, if you defined tools as:

[
  {
    "name": "get_stock_price",
    "description": "Get the current stock price for a given ticker symbol.",
    "input_schema": {
      "type": "object",
      "properties": {
        "ticker": {
          "type": "string",
          "description": "The stock ticker symbol, e.g. AAPL for Apple Inc."
        }
      },
      "required": ["ticker"]
    }
  }
]

And then asked the model "What's the S&P 500 at today?", the model might produce tool_use content blocks in the response like this:

[
  {
    "type": "tool_use",
    "id": "toolu_01D7FLrfh4GYq7yT1ULFeyMV",
    "name": "get_stock_price",
    "input": { "ticker": "^GSPC" }
  }
]

You might then run your get_stock_price tool with {"ticker": "^GSPC"} as an input, and return the following back to the model in a subsequent user message:

[
  {
    "type": "tool_result",
    "tool_use_id": "toolu_01D7FLrfh4GYq7yT1ULFeyMV",
    "content": "259.75 USD"
  }
]

Tools can be used for workflows that include running client-side tools and functions, or more generally whenever you want the model to produce a particular JSON structure of output.

See our guide for more details.

One of the following:

Tool object{ type, input_schema, name, 7 more }

ToolBash20250124 object{ type: "bash_20250124", name, allowed_callers, 4 more }

CodeExecutionTool20250522 object{ type: "code_execution_20250522", name, allowed_callers, 3 more }

CodeExecutionTool20250825 object{ type: "code_execution_20250825", name, allowed_callers, 3 more }

CodeExecutionTool20260120 object{ type: "code_execution_20260120", name, allowed_callers, 3 more }

A tools entry for the code execution tool, version code_execution_20260120. It adds REPL state persistence to version code_execution_20250825.

REPL state persistence applies with programmatic tool calling. With it, the Python interpreter state, such as variable bindings, persists across requests that reuse the container.

See the code execution tool guide to learn what each tool version adds.

CodeExecutionTool20260521 object{ type: "code_execution_20260521", name, allowed_callers, 3 more }

Code execution tool with REPL state persistence.

BrowserToolset20260801 object{ type: "browser_toolset_20260801", cache_control, configs }

The browser toolset: a single tools[] entry (carrying no name) that declares the browser tool family. The model is served the family's tool with any members disabled via configs removed from its schema.

MemoryTool20250818 object{ type: "memory_20250818", name, allowed_callers, 4 more }

ComputerToolset20260801 object{ type: "computer_toolset_20260801", cache_control, configs }

The computer toolset: a single tools[] entry (carrying no name) that declares the computer tool family. The model is served the family's tool with any members disabled via configs removed from its schema. Every member is enabled by default, zoom included. The single-tool options display_number and enable_zoom are not fields of a toolset entry — it carries only type, configs, and cache_control; zoom is controlled via configs.zoom.enabled.

ToolTextEditor20250124 object{ type: "text_editor_20250124", name, allowed_callers, 4 more }

ToolTextEditor20250429 object{ type: "text_editor_20250429", name, allowed_callers, 4 more }

ToolTextEditor20250728 object{ type: "text_editor_20250728", name, allowed_callers, 5 more }

WebSearchTool20250305 object{ type: "web_search_20250305", name, allowed_callers, 7 more }

WebFetchTool20250910 object{ type: "web_fetch_20250910", name, allowed_callers, 9 more }

WebSearchTool20260209 object{ type: "web_search_20260209", name, allowed_callers, 7 more }

WebFetchTool20260209 object{ type: "web_fetch_20260209", name, allowed_callers, 9 more }

WebFetchTool20260309 object{ type: "web_fetch_20260309", name, allowed_callers, 10 more }

Web fetch tool with use_cache parameter for bypassing cached content.

WebSearchTool20260318 object{ type: "web_search_20260318", name, allowed_callers, 8 more }

WebFetchTool20260318 object{ type: "web_fetch_20260318", name, allowed_callers, 11 more }

ToolSearchToolBm25_20251119 object{ type, name, allowed_callers, 3 more }

ToolSearchToolRegex20251119 object{ type, name, allowed_callers, 3 more }

Este endpoint devuelve un invalid_request_error para algunas entradas que la Messages API sí acepta:

Herramientas de servidor como búsqueda web, obtención web, ejecución de código y búsqueda de herramientas (todas las herramientas de servidor excepto la herramienta de asesor)
El conector MCP
Bloques image o document con una fuente url o file

Envía las imágenes y los PDFs en base64 para contarlos. Para las solicitudes que usan herramientas de servidor o servidores MCP, la respuesta de la Messages API informa los tokens utilizados en su objeto usageVista previa disponible con enlaces a la referencia de la API — presiona Tab hasta que el foco entre a la tarjeta y Escape para volver..

Modelos compatibles

Todos los modelos activos admiten el conteo de tokens.

Contar tokens en mensajes básicos

client = anthropic.Anthropic()

response = client.messages.count_tokens(
    model="claude-opus-5-5",
    system="You are a scientist",
    messages=[{"role": "user", "content": "Hello, Claude"}],
)

print(response.json())

Output

{ "input_tokens": 14 }

Contar tokens en mensajes con herramientas

client = anthropic.Anthropic()

response = client.messages.count_tokens(
    model="claude-opus-5-5",
    tools=[
        {
            "name": "get_weather",
            "description": "Get the current weather in a given location",
            "input_schema": {
                "type": "object",
                "properties": {
                    "location": {
                        "type": "string",
                        "description": "The city and state, e.g. San Francisco, CA",
                    }
                },
                "required": ["location"],
            },
        }
    ],
    messages=[{"role": "user", "content": "What's the weather like in San Francisco?"}],
)

print(response.json())

Output

{ "input_tokens": 403 }

Contar tokens en mensajes con imágenes

import base64
import httpx2

image_url = "https://platform.claude.com/docs/images/vision-example.jpg"
image_media_type = "image/jpeg"
image_data = base64.standard_b64encode(httpx2.get(image_url).content).decode("utf-8")

client = anthropic.Anthropic()

response = client.messages.count_tokens(
    model="claude-opus-5-5",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "source": {
                        "type": "base64",
                        "media_type": image_media_type,
                        "data": image_data,
                    },
                },
                {"type": "text", "text": "Describe this image"},
            ],
        }
    ],
)
print(response.json())

Output

{ "input_tokens": 1028 }

Un bloque de imagen incrustado que establece "oversized_image": "error" se rechaza en el momento del conteo exactamente como lo rechazaría la Messages API.

Contar tokens en mensajes con pensamiento

client = anthropic.Anthropic()

response = client.messages.count_tokens(
    model="claude-opus-5-5",
    thinking={"type": "adaptive"},
    messages=[
        {
            "role": "user",
            "content": "Are there an infinite number of prime numbers such that n mod 4 == 3?",
        },
        {
            "role": "assistant",
            "content": [
                {
                    "type": "thinking",
                    "thinking": "This is a nice number theory question. Let's think about it step by step...",
                    "signature": "EuYBCkQYAiJAgCs1le6/Pol5Z4/JMomVOouGrWdhYNsH3ukzUECbB6iWrSQtsQuRHJID6lWV...",
                },
                {
                    "type": "text",
                    "text": "Yes, there are infinitely many prime numbers p such that p mod 4 = 3...",
                },
            ],
        },
        {"role": "user", "content": "Can you write a formal proof?"},
    ],
)

print(response.json())

Output

{ "input_tokens": 88 }

Contar tokens en mensajes con PDFs

import base64
import anthropic

client = anthropic.Anthropic()

with open("/path/to/document.pdf", "rb") as pdf_file:
    pdf_base64 = base64.standard_b64encode(pdf_file.read()).decode("utf-8")

response = client.messages.count_tokens(
    model="claude-opus-5-5",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "document",
                    "source": {
                        "type": "base64",
                        "media_type": "application/pdf",
                        "data": pdf_base64,
                    },
                },
                {"type": "text", "text": "Please summarize this document."},
            ],
        }
    ],
)

print(response.json())

Output

{ "input_tokens": 2188 }

Conteos de tokens en los modelos Claude Fable y Claude Mythos

Claude Fable 5.2, Claude Mythos 5.2, Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5 y Claude Mythos 5 comparten el tokenizador introducido con Claude Opus 4.7. Un prompt cuenta lo mismo en los seis, y aproximadamente un 30 por ciento más que en los modelos anteriores a Claude Opus 4.7 (el aumento exacto depende del contenido). El endpoint de conteo de tokens cuenta según el tokenizador del model que pases. Para medir la diferencia en tu carga de trabajo, cuenta la misma solicitud una vez con tu modelo actual y otra con el modelo al que planeas migrar. Luego compara los dos valores de input_tokens.

Precios y límites de velocidad

El conteo de tokens es de uso gratuito, pero está sujeto a límites de velocidad de solicitudes por minuto según tu nivel de uso. Si necesitas límites más altos, usa Solicitar aumento de nivel en la página Límites de frecuencia de la Claude Console.

Nivel de uso	Solicitudes por minuto (RPM)
Start	5,000
Build	10,000
Scale	20,000

Preguntas frecuentes

Próximos pasos

Contar tokens de mensajes

Lee la referencia completa de la API para el endpoint de conteo de tokens.

Ventanas de contexto

Usa los conteos de tokens para mantener los prompts dentro de la ventana de contexto de un modelo.

Límites de velocidad

Verifica los conteos de tokens antes de enviar una solicitud para mantenerte dentro de tu nivel de uso.

Almacenamiento en caché de prompts

Reduce el costo y la latencia en prompts repetidos almacenando en caché los prefijos de los prompts.

Disponibilidad

Conteo de tokens

API: Messages API
Managed Agents No compatible
Modelos: Todos los modelos

Manejo de datos

Descripción general del manejo de datos

Endpoints aplicables

/v1/messages/count_tokens

Retención de datos ceroElegible

Excluye Modelos cubiertos

Cuenta los tokens antes de enviar solicitudes.

HIPAAElegible

Cómo contar los tokens de un mensaje

Count tokens in a Message

Encabezados anthropic-version anthropic-user-profile-id anthropic-workspace-id

Cuerpo messages model cache_control output_config speed system +3 más

Valor devuelto input_tokens

Modelos compatibles

Contar tokens en mensajes básicos

Contar tokens en mensajes con herramientas

Contar tokens en mensajes con imágenes

Contar tokens en mensajes con pensamiento

Contar tokens en mensajes con PDFs

Conteos de tokens en los modelos Claude Fable y Claude Mythos

Precios y límites de velocidad

Preguntas frecuentes

¿El conteo de tokens usa el almacenamiento en caché de prompts?

Próximos pasos

Disponibilidad

Manejo de datos

Count tokens in a Message

Encabezados anthropic-version anthropic-user-profile-id anthropic-workspace-id

Cuerpo messages model cache_control output_config speed system +3 más

Valor devuelto input_tokens