Mouse coding agent achieved 23 of 30 passes on FrontierHarness tasks using GLM-5.3-Flash at $0.29 per pass, performing comparably to other agents on the full GLM-5.3 model while using significantly cheaper inference. Mouse missed two tasks that Kimi K3 passed on DeepSWE benchmarks but matched K3's performance on other shared failures.
Mouse, built on OpenCode, achieved first place on the FrontierHarness benchmark with a score of 25/30, up from OpenCode's last-place 15/30. The key improvement was a verification loop that enforces a fixed sequence of deterministic steps—checking git changes, running repo checks, verifying no tests were deleted, and validating a MOUSE_AUDIT file—allowing the agent to iteratively refine solutions until all requirements pass, achieving an impressive 90.6% cache hit rate.