HarnessOpt-Bench - Terminal-Bench: leaderboard
Metric: Normalized gain ((held-out score minus seed score) / (1 minus seed score); negative when the nominated harness is worse than the seed; mean over the contestant's optimization rounds, each held-out case scored three times; Terminal-Bench 2.0 terminal tasks with a grok-build target agent; seed held-out score 0.241, 36 test cases; the optimizer edits the seed harness under fixed development and validation evaluation budgets and never sees the test partition). Source: arxiv.org. Saturation forecast: Around May 2028. 10 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 5 | 0.29 |
| 2 | Claude Sonnet 5 | 0.15 |
| 3 | GPT-5.6 Sol | 0.13 |
| 4 | Kimi K3 | 0.12 |
| 5 | GPT-5.6 Terra | 0.04 |
Interactive version: theaggregate.ai/benchmark?slug=harnessopt-bench-terminal-bench · How It Works · Data refreshed daily, snapshot 2026-09-29.