RunUsage reports 0 tokens for Anthropic Models

[ifoukarakis]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

Issue: RunUsage reports 0 tokens while actual token counts appear in details field.

When creating an agent using Anthropic's models and executing it with a query, the RunUsage object in the result incorrectly reports 0 tokens for all token metrics. However, the actual token counts are present and accessible in the details field.

Message history contains the expected values is ModelResponse object.

Expected behavior: Token usage should be properly populated in the RunUsage object's top-level fields.

Example Code

from pydantic_ai import Agent

def main():
    # Create a simple agent with Anthropic model
    agent = Agent("anthropic:claude-3-7-sonnet-20250219")
    # agent = Agent('openai:gpt-5')
    query = "What is 2+2?"

    result = agent.run_sync(query)
    usage = result.usage()
    print(f"Query: {query}")
    print(f"Response: {result.output}")
    print(f"  Requests: {usage.requests}")
    print(f"  Input tokens: {usage.input_tokens}")
    print(f"  Output tokens: {usage.output_tokens}")
    print(f"  Details: {usage.details}")
    for message in result.all_messages():
        print(message)

if __name__ == "__main__":
    main()


=== Output ===

Query: What is 2+2?
Response: 2+2 = 4
  Requests: 1
  Input tokens: 0
  Output tokens: 0
  Details: {'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0, 'input_tokens': 14, 'output_tokens': 11}
ModelRequest(parts=[UserPromptPart(content='What is 2+2?', timestamp=datetime.datetime(2025, 10, 22, 14, 12, 56, 814250, tzinfo=datetime.timezone.utc))])
ModelResponse(parts=[TextPart(content='2+2 = 4')], usage=RequestUsage(details={'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0, 'input_tokens': 14, 'output_tokens': 11}), model_name='claude-3-7-sonnet-20250219', timestamp=datetime.datetime(2025, 10, 22, 14, 12, 58, 10504, tzinfo=datetime.timezone.utc), provider_name='anthropic', provider_details={'finish_reason': 'end_turn'}, provider_response_id='msg_01BGULjbBK76JZCqno7uqAXb', finish_reason='stop')

Python, Pydantic AI & LLM client version

Python 3.10.16

anthropic==0.71.0
pydantic-ai==1.2.0

[DouweM]

@alexmojaki Can you have a look here please? You recently did some refactoring here: #3115.

[ifoukarakis]

Thank you!