Anthropic Python SDK v0.104.0 adds thinking token count estimates for streaming responses

Anthropic released version 0.104.0 of its Python SDK on May 21, 2026, adding support for tracking token usage in reasoning model thought processes.

What's new

The update introduces a thinking-token-count beta feature that provides estimated token counts within thinking block deltas during streaming responses. This allows developers to monitor token consumption in real-time as Claude's reasoning models process extended chains of thought.

The feature specifically targets streaming scenarios where Claude models with thinking capabilities—such as Claude 3.5 Sonnet with extended thinking mode—generate internal reasoning before producing final outputs.

Technical details

The implementation provides token count estimates as part of the delta stream, enabling developers to:

• Track thinking token usage during active streaming
• Estimate costs for reasoning operations in real-time
• Monitor and debug extended thinking processes
• Optimize prompts based on thinking token consumption

The feature is marked as beta, indicating the API may change in future releases.

Version information

• Version: 0.104.0
• Release date: May 21, 2026
• Type: Minor version update
• Full changelog: Available at github.com/anthropics/anthropic-sdk-python/compare/v0.103.1...v0.104.0

What this means

This update addresses a key observability gap for developers using Claude's reasoning capabilities. Previously, tracking token usage in thinking blocks required waiting for complete responses. With streaming token counts, developers can now monitor costs and performance in real-time, particularly important for applications using extended thinking modes where reasoning token counts can significantly exceed output tokens. The beta designation suggests Anthropic is still refining how thinking token metrics are calculated and surfaced to developers.