assafelovic--gpt-researcher/PR_pr_feat-anthropic-real-usage-cost-tracking.md

One-line summary

Track Anthropic costs from native usage metadata instead of relying on the fallback tiktoken estimator.

Problem

Anthropic requests were being costed with a generic text-length estimator rather than the provider's native token usage metadata. That made Anthropic totals inaccurate, especially for streaming responses and tool-calling flows where usage arrives through response metadata rather than plain text length.

Solution

• add Anthropic-specific pricing rules in gpt_researcher/utils/costs.py
• prefer response_metadata["usage"] / usage_metadata when Anthropic returns native token counts
• capture usage and response metadata in GenericLLMProvider for both normal and streaming responses
• use the same cost-calculation path in create_chat_completion() and tool-calling flows
• add focused tests for pricing and streaming metadata capture

Testing

Verified locally with:

• uv run python -c "import gpt_researcher; print('ok')"
• uv run python -m unittest tests.test_costs tests.test_llm_usage_tracking

Breaking changes

None.

Related issues

None identified.