Anthropic streaming spans keep message_start input/cache token counts instead of the final message_delta usage #25621

Is there an existing issue for this?

• I have checked for existing issues
• I have reviewed the documentation
• I am using the latest SDK release

How do you use Sentry?

Sentry SaaS (sentry.io)

Which SDK are you using?

@sentry/node

SDK Version

11.8.0 (also develop at 6fb6b07)

Framework Version

Node 24, @anthropic-ai/sdk with stream: true

Link to Sentry event

No response

Reproduction Example/SDK Setup

Sentry.init({ integrations: [Sentry.anthropicAIIntegration()], tracesSampleRate: 1 });
const stream = client.messages.stream({ model, max_tokens: 1024, tools: [{ type: 'web_search_20250305', name: 'web_search' }], messages });
for await (const _ of stream) {}

Steps to Reproduce

1. Stream an Anthropic message whose message_delta.usage reports larger cumulative input_tokens / cache_creation_input_tokens / cache_read_input_tokens than message_start (server tools such as web search do this; a mocked stream with message_start input 50 / cache 10 / 20 and a final message_delta with input 120 / cache 30 / 70 shows it too).
2. Look at the gen_ai.usage.input_tokens and cache attributes of the resulting span.

Expected Result

The span reports the final cumulative counts from message_delta (120 / 30 / 70 in the mocked stream), since Anthropic's message_delta usage is the running total.

Actual Result

The span keeps the message_start values (50 / 10 / 20). handleMessageMetadata in packages/server-utils/src/ai/anthropic-ai/streaming.ts reads only output_tokens from message_delta.

Additional Context

Proposed fix: when message_delta.usage carries numeric input_tokens / cache_* values, overwrite the message_start values with them (they are totals, not increments) and recompute the prompt total; keep the message_start values when the delta omits them. A patch with tests is ready; happy to open the PR.