Anthropic/MiniMax streaming: reported task cost omits output-token cost #1727

Summary

For Anthropic and MiniMax models, the task cost reported in the UI omits the output-token portion of the bill. Every task on the default Anthropic provider shows only input + cache-read/write cost — typically 20–40% of the real total.

Mechanism

In src/api/providers/anthropic.ts the streaming loop accumulates outputTokens only from message_start.usage.output_tokens, which Anthropic documents (and the pinned SDK @anthropic-ai/[email protected] MessageStream confirms, lib/MessageStream.js:454 — replace semantics) as a seed value (0 or 1). The real total arrives cumulatively in message_delta.usage.output_tokens, whose final event carries the full count. The message_delta branch yields the value to the consumer but never adopts it into the local accumulator, so the end-of-stream totalCost chunk is computed from the seed.

Task.ts then assigns totalCost = chunk.totalCost (overriding the consumer-side costResult which is summed correctly), so the wrong value wins. Same defect in src/api/providers/minimax.ts. In-repo proof this is a known trap: opencode-go.ts:558-565 fixes exactly this with an explanatory comment.

Repro

1. Select any Anthropic (or MiniMax) model, run any task.
2. Open the task cost/token summary.
3. Compare reported cost against calculateApiCostAnthropic(modelInfo, inputTokens, outputTokens) with the real output count — the output-token term (60–80% of the bill) is missing. Token counts look right because the consumer sums the deltas itself.

Suggested fix

Adopt (not add) the cumulative value in the message_delta case: outputTokens = chunk.usage.output_tokens || outputTokens, mirrored in minimax.ts. Regression tests: stream with message_start usage {input:100, output:1} + message_delta usage {output_tokens:500} → final usage chunk's totalCost must equal cost(info, 100, 500).