EdgeBench: leaderboard
EdgeBench official published mean best-so-far score after 12 hours over 134 task environments. Harness and compaction context are separate; the unsuffixed board uses Claude Code with 200k compaction. Only valid trajectories are scored under the primary missing-case policy; no animation-curve mean is substituted. Board: EdgeBench.
Metric: Published score @12h (134 tasks). Source: edge-bench.org. Saturation forecast: Around November 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.1 | 37.4 | #97 |
| 2 | DeepSeek V4 Pro Preview | 31 | #148 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=edgebench · How It Works · Data refreshed daily, snapshot 2026-10-20.