Token Counting

​

What is counted

• Messages
  • Content of the message - Includes the system message, user message, and the assistant message content.
  • Tool call arguments and results
  • Optional reasoning content
  • Multimodal content blocks
• Tools
  • Tool definitions can be a meaningful part of the total token count, especially with large parameter schemas or long descriptions.
• Output schema
  • If you use an Output Schema, the schema is included in the token count.
• Multimodal attachments
  • Images, audio, video, and files attached to messages are counted using conservative estimates.

​

Optional dependencies (recommended)

uv pip install -U openai tiktoken tokenizers

• openai: model provider used in the example below.
• tiktoken: used when available for OpenAI-style tokenization.
• tokenizers: used for certain open-source tokenizers when available.
• If neither is available for the given model, we fall back to heuristic estimates.

​

Example: counting tokens

from pydantic import BaseModel

from agno.models.message import Message
from agno.models.openai import OpenAIResponses


class Answer(BaseModel):
    answer: str


model = OpenAIResponses(id="gpt-5.2")

messages = [
    Message(role="system", content="You are a concise assistant."),
    Message(role="user", content="Summarize context compression in 2 sentences."),
]

# Tool definitions can be passed as OpenAI-style tool dicts
tools = [
    {
        "type": "function",
        "function": {
            "name": "search_web",
            "description": "Search the web for a query.",
            "parameters": {
                "type": "object",
                "properties": {"query": {"type": "string"}},
                "required": ["query"],
            },
        },
    }
]

tokens = model.count_tokens(messages=messages, tools=tools, output_schema=Answer)
print(f"Estimated tokens: {tokens}")

​

Token counting in token-based context compression

​

Multimodal estimates

• Images: estimated via a tile-based approach (vision-style counting)
• Audio: estimated using tokens-per-second
• Video: estimated as frames counted similarly to images (with conservative defaults if fps/dimensions are unknown)
• Files: estimated based on file type/size

​

Notes

• Token counting is still in Beta. We do our best to provide an estimate but we do not claim it to be 100% accurate across all providers and models. Please be wary of using this token count for calculating costs.
• For some providers like Claude we are able to call the exact endpoint and get an exact token count. This is not supported for all providers yet.