HERA-Bench - Final Harness - Paired Success: leaderboard
HERA-Bench (2026): explicit Appendix A1.T4 reasoning and routed precision retained per model; 60 matched feasible/infeasible pairs, 120 tasks with executable verifiers. The final harness is evolved using GPT-5.6 Luna and fixed for cross-model evaluation; the base harness is separate. Paired success requires correct action and correct abstention on the same pair; no training-pool measurements mixed in. Rate (%); higher is better.
Metric: Rate (%). Source: arxiv.org. Saturation forecast: Around January 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (High) | 80 | #6 (GPT-6 Astra) |
| 2 | GPT-5.6 Sol (High) | 78.3 | #18 (GPT-5.6 Sol) |
| 3 | GPT-5.6 Terra (High) | 73.3 | #41 (GPT-5.6 Terra) |
| 4 | GPT-5.6 Luna | 70 | #70 |
| 5 | Grok 4.6 (High) | 66.7 | #25 (Grok 4.6) |
| 6 | Claude Opus 5 (High) | 61.7 | #12 (Claude Opus 5) |
| 7 | Gemini 3.5 Flash Lite (High) | 61.7 | #157 (Gemini 3.5 Flash Lite) |
| 8 | Gemini 3.8 Flash (High) | 50 | #23 (Gemini 3.8 Flash) |
| 9 | Gemini 3.1 Pro (Preview) (High) | 48.3 | #58 (Gemini 3.1 Pro (Preview)) |
| 10 | DeepSeek V4.1 Flash (High) | 46.7 | #36 (DeepSeek V4.1 Flash) |
| 11 | Claude Sonnet 5 (High) | 41.7 | #56 (Claude Sonnet 5) |
| 12 | Mistral Medium 3.5 (High) | 20 | #357 (Mistral Medium 3.5) |
| 13 | Mistral Large 3 | 5 | #402 |
No result here: #4 Claude Opus 5.5, #7 Gemini 4 Argon, #8 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=hera-bench-final-harness-paired-success · How It Works · Data refreshed daily, snapshot 2026-10-21.