CLAIM-2026-08-11-08

On WeaveBench with Qwen 3.7-Plus using Claude Code, LongHorizon-Harness raised pass rate from a 51.8% baseline to 80.7%; on Terminal-Bench 2.1 with Qwen 3.7-Plus, success rate rose from 69.7% to 77.2%.

Why it matters

Gives the argument a number. Attribution to benchmark and model matters — it is not a general result.

Sources

• arxiv2026longhoriz