HarnessOpt-Bench - GAIA: leaderboard
The GAIA task of HarnessOpt-Bench (Scale AI, 2026): an optimizer LLM working through a coding harness must turn a non-functional stub into a working general-assistant agent under fixed evaluation budgets. Because the stub scores zero, the normalized gain equals the held-out GAIA score of the built agent, from 0 to 1; higher is better.
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; GAIA assistant questions with a gpt-5.4-mini target agent; the seed is a non-functional stub scoring 0, so the gain equals the built agent's held-out score, 66 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 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 | 0.47 | #22 |
| 2 | Kimi K3 | 0.34 | #54 |
| 3 | GPT-5.6 Sol | 0.31 | #30 |
| 4 | Claude Sonnet 5 | 0.25 | #111 |
| 5 | GPT-5.6 Terra | 0.17 | #62 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=harnessopt-bench-gaia · How It Works · Data refreshed daily, snapshot 2026-11-13.