stream.get_final_message() does not return the correct usage of output_tokens

TL;DR

Contributor RobertCraigie: the current behavior is correct - the final messagedelta usage already equals the total, identical to a non-streaming request's usage; the initial messagestart usage can be ignored. Reporter confirmed understanding.

The error

Issue anthropics/anthropic-sdk-python#424 (closed, 9 comments): After streaming, stream.get_final_message() usage showed output_tokens that did not match expectations - the reporter expected to sum message_start and message_delta usage.

Fix

1. Contributor RobertCraigie: the current behavior is correct - the final messagedelta usage already equals the total, identical to a non-streaming request's usage; the initial messagestart usage can be ignored.
2. Reporter confirmed understanding.

Success check: the symptom above is gone and the original operation completes.

When to use

• You hit this exact symptom in anthropic-sdk-python.
• You want the verified fix without digging through the thread.

When not to use

• A different error in anthropic-sdk-python, even a similar-looking one.
• General anthropic-sdk-python setup or credential questions.

Compatibility

• Anthropic SDK Python (anthropics/anthropic-sdk-python)

Variant phrasings

• How to fix "stream.getfinalmessage() does not return the correct usage of output_tokens" in anthropic-sdk-python
• anthropic-sdk-python: stream.getfinalmessage() does not return the correct usage of output_tokens

Source

• https://github.com/anthropics/anthropic-sdk-python/issues/424

Asked in gh:anthropics/anthropic-sdk-python#424 (https://github.com/anthropics/anthropic-sdk-python/issues/424): stream.getfinalmessage() does not return the correct usage of outputtokens. The issue report: Issue anthropics/anthropic-sdk-python#424 (closed, 9 comments): After streaming, stream.getfinalmessage() usage showed outputtokens that did not match expectations - the reporter expected to sum messagestart and messagedelta usage. The verified answer: Contributor RobertCraigie: the current behavior is correct - the final messagedelta usage already equals the total, identical to a non-streaming request's usage; the initial messagestart usage can be ignored. Reporter confirmed understanding. Lesson: for streamed Anthropic requests, use the final usage block as the total; do not sum messagestart + messagedelta.