Entender e contar tokens

O Gemini e outros modelos de IA generativa processam entradas e saídas em uma granularidade chamada token.

Para modelos do Gemini, um token equivale a cerca de quatro caracteres. 100 tokens equivalem a cerca de 60 a 80 palavras em inglês.

Sobre tokens

Os tokens podem ser caracteres únicos, como z, ou palavras inteiras, como cat. Palavras longas são divididas em vários tokens. O conjunto de todos os tokens usados pelo modelo é chamado de vocabulário, e o processo de dividir o texto em tokens é chamado de tokenização.

Quando o faturamento está ativado, o custo de uma chamada para a API Gemini é determinado em parte pelo número de tokens de entrada e saída. Por isso, saber como contar tokens pode ser útil.

Contar tokens

Todas as entradas e saídas da API Gemini são tokenizadas, incluindo texto, arquivos de imagem e outras modalidades que não são de texto.

É possível contar tokens das seguintes maneiras:

  • Chame count_tokens com a entrada da solicitação. Retorna o número total de tokens apenas na entrada. Faça essa chamada antes de enviar a entrada para verificar o tamanho das suas solicitações.

  • Use o usage na resposta da interação. Retorna contagens de tokens para entrada (total_input_tokens), saída (total_output_tokens), pensamento (total_thought_tokens), conteúdo em cache (total_cached_tokens), uso de ferramentas (total_tool_use_tokens) e total (total_tokens).

Contar tokens de texto

Python

# This will only work for SDK newer than 2.0.0
from google import genai

client = genai.Client()
prompt = "The quick brown fox jumps over the lazy dog."

# Count tokens before sending
total_tokens = client.models.count_tokens(
    model="gemini-3.8-flash",
    contents=prompt
)
print("total_tokens:", total_tokens.total_tokens)

# Get usage from interaction
interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input=prompt
)
print(interaction.usage)

JavaScript

// This will only work for SDK newer than 2.0.0
import { GoogleGenAI } from '@google/genai';

const client = new GoogleGenAI({});
const prompt = "The quick brown fox jumps over the lazy dog.";

// Count tokens before sending
const countResponse = await client.models.countTokens({
    model: "gemini-3.8-flash",
    contents: prompt,
});
console.log(countResponse.totalTokens);

// Get usage from interaction
const interaction = await client.interactions.create({
    model: "gemini-3.8-flash",
    input: prompt,
});
console.log(interaction.usage);

Java

import com.google.genai.Client;
import com.google.genai.gaos.models.interactions.CreateModelInteraction;
import com.google.genai.gaos.models.interactions.Interaction;
import com.google.genai.gaos.models.interactions.InteractionsInput;
import com.google.genai.gaos.models.interactions.Model;
import com.google.genai.gaos.models.operations.CreateInteractionRequestBody;
import com.google.genai.types.CountTokensResponse;

Client client = new Client();
String prompt = "The quick brown fox jumps over the lazy dog.";

// Count tokens before sending
CountTokensResponse countResponse =
    client.models.countTokens("gemini-3.8-flash", prompt, null);
System.out.println("total_tokens: " + countResponse.totalTokens().orElse(0));

// Get usage from interaction
CreateModelInteraction params =
    CreateModelInteraction.builder()
        .model(Model.of("gemini-3.8-flash"))
        .input(InteractionsInput.of(prompt))
        .build();

Interaction interaction =
    client.interactions.create(CreateInteractionRequestBody.of(params)).interaction().get();
System.out.println(interaction.usage().orElse(null));

Go

package main

import (
    "context"
    "fmt"
    "log"

    "google.golang.org/genai"
)

func main() {
    ctx := context.Background()
    client, err := genai.NewClient(ctx, nil)
    if err != nil {
        log.Fatal(err)
    }

    modelInfo, err := client.Models.Get(ctx, "gemini-3.8-flash", nil)
    if err != nil {
        log.Fatal(err)
    }

    fmt.Printf("Input token limit: %d\n", modelInfo.InputTokenLimit)
    fmt.Printf("Output token limit: %d\n", modelInfo.OutputTokenLimit)
}

REST

# Specifies the API revision to avoid breaking changes when they become default
curl -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.8-flash:countTokens" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents": [{"parts": [{"text": "The quick brown fox."}]}]}'

Contar tokens multiturno

Contar tokens no histórico de conversas usando previous_interaction_id:

Python

# This will only work for SDK newer than 2.0.0
# First interaction
interaction1 = client.interactions.create(
    model="gemini-3.8-flash",
    input="Hi, my name is Bob"
)

# Second interaction continues the conversation
interaction2 = client.interactions.create(
    model="gemini-3.8-flash",
    input="What's my name?",
    previous_interaction_id=interaction1.id
)

# Usage includes tokens from both turns
print(f"Input tokens: {interaction2.usage.total_input_tokens}")
print(f"Output tokens: {interaction2.usage.total_output_tokens}")
print(f"Total tokens: {interaction2.usage.total_tokens}")

JavaScript

// This will only work for SDK newer than 2.0.0
// First interaction
const interaction1 = await client.interactions.create({
    model: "gemini-3.8-flash",
    input: "Hi, my name is Bob"
});

// Second interaction continues the conversation
const interaction2 = await client.interactions.create({
    model: "gemini-3.8-flash",
    input: "What's my name?",
    previous_interaction_id: interaction1.id
});

console.log(`Input tokens: ${interaction2.usage.total_input_tokens}`);
console.log(`Output tokens: ${interaction2.usage.total_output_tokens}`);

Java

import com.google.genai.Client;
import com.google.genai.gaos.models.interactions.CreateModelInteraction;
import com.google.genai.gaos.models.interactions.Interaction;
import com.google.genai.gaos.models.interactions.InteractionsInput;
import com.google.genai.gaos.models.interactions.Model;
import com.google.genai.gaos.models.interactions.Usage;
import com.google.genai.gaos.models.operations.CreateInteractionRequestBody;

Client client = new Client();

// First interaction
CreateModelInteraction params1 =
    CreateModelInteraction.builder()
        .model(Model.of("gemini-3.8-flash"))
        .input(InteractionsInput.of("Hi, my name is Bob"))
        .build();

Interaction interaction1 =
    client.interactions.create(CreateInteractionRequestBody.of(params1)).interaction().get();

// Second interaction continues the conversation
CreateModelInteraction params2 =
    CreateModelInteraction.builder()
        .model(Model.of("gemini-3.8-flash"))
        .input(InteractionsInput.of("What's my name?"))
        .previousInteractionId(interaction1.id().orElse(""))
        .build();

Interaction interaction2 =
    client.interactions.create(CreateInteractionRequestBody.of(params2)).interaction().get();

// Usage includes tokens from both turns
if (interaction2.usage().isPresent()) {
  Usage usage = interaction2.usage().get();
  System.out.println("Input tokens: " + usage.totalInputTokens().orElse(0));
  System.out.println("Output tokens: " + usage.totalOutputTokens().orElse(0));
  System.out.println("Total tokens: " + usage.totalTokens().orElse(0));
}

Go

package main

import (
    "context"
    "fmt"
    "log"

    "google.golang.org/genai"
    "google.golang.org/genai/interactions/models/interactions"
    "google.golang.org/genai/interactions/models/operations"
)

func main() {
    ctx := context.Background()
    client, err := genai.NewClient(ctx, nil)
    if err != nil {
        log.Fatal(err)
    }

    prompt := "The quick brown fox jumps over the lazy dog."

    // Count input tokens before sending
    totalTokens, err := client.Models.CountTokens(ctx, "gemini-3.8-flash", genai.Text(prompt), nil)
    if err != nil {
        log.Fatal(err)
    }
    fmt.Printf("total_tokens: %d\n", totalTokens.TotalTokens)

    // Create the interaction and inspect the returned usage metadata
    res, err := client.Interactions.Create(ctx, operations.CreateInteractionRequest{
        Body: operations.NewCreateInteractionRequestBody(interactions.CreateModelInteraction{
            Model: interactions.Model("gemini-3.8-flash"),
            Input: interactions.NewInteractionsInput(prompt),
        }),
    })
    if err != nil {
        log.Fatal(err)
    }

    interaction := res.Interaction
    if interaction.OutputText != nil {
        fmt.Println(*interaction.OutputText)
    }
    if interaction.Usage != nil {
        if interaction.Usage.TotalInputTokens != nil {
            fmt.Printf("Input tokens: %d\n", *interaction.Usage.TotalInputTokens)
        }
        if interaction.Usage.TotalOutputTokens != nil {
            fmt.Printf("Output tokens: %d\n", *interaction.Usage.TotalOutputTokens)
        }
        if interaction.Usage.TotalThoughtTokens != nil {
            fmt.Printf("Thought tokens: %d\n", *interaction.Usage.TotalThoughtTokens)
        }
        if interaction.Usage.TotalTokens != nil {
            fmt.Printf("Total tokens: %d\n", *interaction.Usage.TotalTokens)
        }
    }
}

Contar tokens multimodais

Todas as entradas da API Gemini são tokenizadas, incluindo imagens, vídeos e áudios. Pontos principais sobre a tokenização:

  • Imagens: imagens ≤384 pixels em ambas as dimensões contam como 258 tokens. Imagens maiores são divididas em blocos de 768 x 768 pixels, cada um contando como 258 tokens.
  • Vídeo: 263 tokens por segundo (aplicável ao processamento estático). Para o processamento agêntico, o uso de tokens varia. Consulte Uso de tokens de vídeo por modo de processamento.
  • Áudio: 32 tokens por segundo

Tokens de imagem

Python

# This will only work for SDK newer than 2.0.0
uploaded_file = client.files.upload(file="path/to/image.jpg")

# Count tokens for image + text
total_tokens = client.models.count_tokens(
    model="gemini-3.8-flash",
    contents=["Tell me about this image", uploaded_file]
)
print(f"Total tokens: {total_tokens}")

# Generate with image
interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input=[
        {"type": "text", "text": "Tell me about this image"},
        {"type": "image", "uri": uploaded_file.uri, "mime_type": uploaded_file.mime_type}
    ]
)
print(interaction.usage)

JavaScript

// This will only work for SDK newer than 2.0.0
const uploadedFile = await client.files.upload({
    file: "path/to/image.jpg",
    config: { mimeType: "image/jpeg" }
});

// Count tokens
const countResponse = await client.models.countTokens({
    model: "gemini-3.8-flash",
    contents: [
        { text: "Tell me about this image" },
        { fileData: { fileUri: uploadedFile.uri, mimeType: uploadedFile.mimeType } }
    ]
});
console.log(countResponse.totalTokens);

Java

import com.google.genai.Client;
import com.google.genai.gaos.models.interactions.CreateModelInteraction;
import com.google.genai.gaos.models.interactions.ImageContent;
import com.google.genai.gaos.models.interactions.ImageContentMimeType;
import com.google.genai.gaos.models.interactions.Interaction;
import com.google.genai.gaos.models.interactions.InteractionsInput;
import com.google.genai.gaos.models.interactions.Model;
import com.google.genai.gaos.models.interactions.TextContent;
import com.google.genai.gaos.models.operations.CreateInteractionRequestBody;
import com.google.genai.types.Content;
import com.google.genai.types.CountTokensResponse;
import com.google.genai.types.File;
import com.google.genai.types.Part;
import com.google.genai.types.UploadFileConfig;
import java.util.Arrays;

Client client = new Client();

File uploadedFile =
    client.files.upload(
        new java.io.File("path/to/image.jpg"),
        UploadFileConfig.builder().mimeType("image/jpeg").build());

// Count tokens for image + text
CountTokensResponse countResponse =
    client.models.countTokens(
        "gemini-3.8-flash",
        Arrays.asList(
            Content.fromParts(
                Part.fromText("Tell me about this image"),
                Part.fromUri(
                    uploadedFile.uri().orElse(""), uploadedFile.mimeType().orElse("image/jpeg")))),
        null);
System.out.println("Total tokens: " + countResponse.totalTokens().orElse(0));

// Generate with image
CreateModelInteraction params =
    CreateModelInteraction.builder()
        .model(Model.of("gemini-3.8-flash"))
        .input(
            InteractionsInput.ofContent(
                Arrays.asList(
                    TextContent.builder().text("Tell me about this image").build(),
                    ImageContent.builder()
                        .uri(uploadedFile.uri().orElse(""))
                        .mimeType(
                            ImageContentMimeType.of(uploadedFile.mimeType().orElse("image/jpeg")))
                        .build())))
        .build();

Interaction interaction =
    client.interactions.create(CreateInteractionRequestBody.of(params)).interaction().get();
System.out.println(interaction.usage().orElse(null));

Go

package main

import (
    "context"
    "fmt"
    "log"

    "google.golang.org/genai"
    "google.golang.org/genai/interactions/models/interactions"
    "google.golang.org/genai/interactions/models/operations"
)

func main() {
    ctx := context.Background()
    client, err := genai.NewClient(ctx, nil)
    if err != nil {
        log.Fatal(err)
    }

    res, err := client.Interactions.Create(ctx, operations.CreateInteractionRequest{
        Body: operations.NewCreateInteractionRequestBody(interactions.CreateModelInteraction{
            Model:  interactions.Model("gemini-3.8-flash"),
            Input:  interactions.NewInteractionsInput("Explain the history of the internet in 3 paragraphs."),
            Stream: genai.Ptr(true),
        }),
    })
    if err != nil {
        log.Fatal(err)
    }
    stream := res.InteractionSSEStreamEvent
    defer stream.Close()

    for stream.Next() {
        event := stream.Value()
        if stepDelta := event.GetDataStepDelta(); stepDelta != nil {
            if textDelta := stepDelta.GetDeltaText(); textDelta != nil {
                fmt.Print(textDelta.GetText())
            }
        }
        if completed := event.GetDataInteractionCompleted(); completed != nil {
            usage := completed.Interaction.Usage
            if usage != nil && usage.TotalTokens != nil {
                fmt.Printf("\nTotal tokens: %d\n", *usage.TotalTokens)
            }
        }
    }
    if err := stream.Err(); err != nil {
        log.Fatal(err)
    }
}

Exemplo de dados inline:

Python

# This will only work for SDK newer than 2.0.0
import base64

with open('image.jpg', 'rb') as f:
    image_bytes = f.read()

interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input=[
        {"type": "text", "text": "Describe this image"},
        {
            "type": "image",
            "data": base64.b64encode(image_bytes).decode('utf-8'),
            "mime_type": "image/jpeg"
        }
    ]
)
print(interaction.usage)

Tokens de vídeo

Python

# This will only work for SDK newer than 2.0.0
import time

video_file = client.files.upload(file="path/to/video.mp4")

while not video_file.state or video_file.state.name != "ACTIVE":
    print("Processing video...")
    time.sleep(5)
    video_file = client.files.get(name=video_file.name)

# A 60-second video is approximately 100 * 60 = 6,000 tokens
total_tokens = client.models.count_tokens(
    model="gemini-3.8-flash",
    contents=["Summarize this video", video_file]
)
print(f"Total tokens: {total_tokens}")

# Generate with video
interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input=[
        {"type": "text", "text": "Summarize this video"},
        {"type": "video", "uri": video_file.uri, "mime_type": video_file.mime_type}
    ]
)
print(interaction.usage)

Uso de tokens de vídeo por modo de processamento

O uso de tokens para vídeo depende do modo de processamento:

Modo de processamento Cálculo de tokens Uso típico
Estático (padrão) Cerca de 100 tokens/segundo por padrão (baixa resolução) ou 300 tokens/segundo (alta resolução). Todos os frames são amostrados a 1 FPS. Previsível e proporcional à duração do vídeo.
Agêntico Varia de acordo com a complexidade do conteúdo. O modelo carrega apenas a transcrição e/ou os frames e/ou o áudio necessários para responder ao comando. Até 88% menos tokens para conteúdo longo.

Com o processamento de agente, uma palestra de uma hora que usaria cerca de 1,08 milhão de tokens no modo estático pode usar cerca de 108 mil tokens, dependendo do comando e do conteúdo.

Para verificar o uso real de tokens em uma solicitação, inspecione interaction.usage. Os tokens de vídeo generativos são informados nos seguintes campos:

  • Comando inicial (referência de vídeo + comando do usuário): total_input_tokens
  • Pensamento de navegação: total_thought_tokens
  • Transcrição, frames e áudio carregados sob demanda: total_tool_use_tokens
  • Resposta final: total_output_tokens

Tokens de áudio

Python

# This will only work for SDK newer than 2.0.0
audio_file = client.files.upload(file="path/to/audio.mp3")

# A 60-second audio clip is approximately 32 * 60 = 1,920 tokens
total_tokens = client.models.count_tokens(
    model="gemini-3.8-flash",
    contents=["Transcribe this audio", audio_file]
)
print(f"Total tokens: {total_tokens}")

# Generate with audio
interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input=[
        {"type": "text", "text": "Transcribe this audio"},
        {"type": "audio", "uri": audio_file.uri, "mime_type": audio_file.mime_type}
    ]
)
print(interaction.usage)

Contar tokens de instruções do sistema

As instruções do sistema são contabilizadas como parte dos tokens de entrada:

Python

# This will only work for SDK newer than 2.0.0
interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input="Hello!",
    system_instruction="You are a helpful assistant who speaks like a pirate."
)

# system_instruction tokens included in total_input_tokens
print(f"Input tokens: {interaction.usage.total_input_tokens}")

Contar tokens de ferramentas

As ferramentas (funções, execução de código, Pesquisa Google) também são contabilizadas:

Python

# This will only work for SDK newer than 2.0.0
tools = [
    {
        "type": "function",
        "name": "get_weather",
        "description": "Get current weather",
        "parameters": {
            "type": "object",
            "properties": {
                "location": {"type": "string"}
            }
        }
    }
]

interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input="What's the weather in Tokyo?",
    tools=tools
)

print(f"Input tokens: {interaction.usage.total_input_tokens}")
print(f"Tool use tokens: {interaction.usage.total_tool_use_tokens}")

Janela de contexto

Cada modelo do Gemini tem um número máximo de tokens que pode processar. A janela de contexto define o limite combinado de tokens de entrada e saída.

Receber o tamanho da janela de contexto de forma programática

Python

# This will only work for SDK newer than 2.0.0
model_info = client.models.get(model="gemini-3.8-flash")
print(f"Input token limit: {model_info.input_token_limit}")
print(f"Output token limit: {model_info.output_token_limit}")

JavaScript

// This will only work for SDK newer than 2.0.0
const modelInfo = await client.models.get({ model: "gemini-3.8-flash" });
console.log(`Input token limit: ${modelInfo.inputTokenLimit}`);
console.log(`Output token limit: ${modelInfo.outputTokenLimit}`);

Java

import com.google.genai.Client;
import com.google.genai.types.Model;

Client client = new Client();

Model modelInfo = client.models.get("gemini-3.8-flash", null);
System.out.println("Input token limit: " + modelInfo.inputTokenLimit().orElse(0));
System.out.println("Output token limit: " + modelInfo.outputTokenLimit().orElse(0));

Go

package main

import (
    "context"
    "encoding/base64"
    "fmt"
    "log"
    "os"

    "google.golang.org/genai"
    "google.golang.org/genai/interactions/models/interactions"
    "google.golang.org/genai/interactions/models/operations"
)

func main() {
    ctx := context.Background()
    client, err := genai.NewClient(ctx, nil)
    if err != nil {
        log.Fatal(err)
    }

    prompt := "Tell me about this instrument"
    imageBytes, err := os.ReadFile("/path/to/organ.jpg")
    if err != nil {
        log.Fatal(err)
    }
    base64Image := base64.StdEncoding.EncodeToString(imageBytes)

    // Count tokens before creating the interaction
    parts := []*genai.Part{
        genai.NewPartFromText(prompt),
        genai.NewPartFromBytes(imageBytes, "image/jpeg"),
    }
    totalTokens, err := client.Models.CountTokens(ctx, "gemini-3.8-flash", []*genai.Content{
        genai.NewContentFromParts(parts, genai.RoleUser),
    }, nil)
    if err != nil {
        log.Fatal(err)
    }
    fmt.Printf("Estimated input tokens: %d\n", totalTokens.TotalTokens)

    // Create the multimodal interaction and inspect the usage metadata
    res, err := client.Interactions.Create(ctx, operations.CreateInteractionRequest{
        Body: operations.NewCreateInteractionRequestBody(interactions.CreateModelInteraction{
            Model: interactions.Model("gemini-3.8-flash"),
            Input: interactions.NewInteractionsInput([]interactions.Content{
                interactions.NewContent(interactions.TextContent{
                    Text: prompt,
                }),
                interactions.NewContent(interactions.ImageContent{
                    Data:     genai.Ptr(base64Image),
                    MimeType: interactions.ImageContentMimeTypeImageJpeg.ToPointer(),
                }),
            }),
        }),
    })
    if err != nil {
        log.Fatal(err)
    }

    interaction := res.Interaction
    if interaction.OutputText != nil {
        fmt.Println(*interaction.OutputText)
    }
    if interaction.Usage != nil && interaction.Usage.TotalTokens != nil {
        fmt.Printf("Total tokens billed: %d\n", *interaction.Usage.TotalTokens)
    }
}

Encontre os tamanhos da janela de contexto na página Modelos.

A seguir