EVMbench: leaderboard
EVMbench evaluates smart-contract agents without hints. This board is the unweighted mean of the three printed task means, not pooled success over a common denominator: Detect uses 120 vulnerabilities, Patch 45 and Exploit 24. Every source reasoning/scaffold run is retained with its disclosed snapshot; task-level scores remain in provenance.
Metric: Unweighted mean of Detect/Patch/Exploit task means (%); no hints. Source: github.com. Saturation forecast: Around April 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.3 Codex (xHigh) | 51 | #69 (GPT-5.3 Codex) |
| 2 | GPT-5.3 Codex (High) | 47.1 | #69 (GPT-5.3 Codex) |
| 3 | GPT-5.2 (2025-12-11) (xHigh) | 47 | |
| 4 | Claude Opus 4.6 | 44.2 | #63 |
| 5 | GPT-5.2 (2025-12-11) (High) | 41.4 | |
| 6 | GPT-5.2 (2025-12-11) (Medium) | 39.1 | |
| 7 | GPT-5.2 | 39 | #110 |
| 8 | Claude Opus 4.5 (20251101) | 36.2 | #74 |
| 9 | GPT-5 | 25.1 | #98 |
| 10 | Gemini 3 Pro (Preview) | 22.4 | #66 |
| 11 | O3 | 14.5 | #124 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=evmbench · How It Works · Data refreshed daily, snapshot 2026-10-20.