SkillEvolBench (SelfGen-Always) - Deployment Success: leaderboard
Skill reuse by coding agents on SkillEvolBench (Lei et al., 2026): 180 verifier-scored tasks in six environments such as code repair, API orchestration and document transformation, each model running in its vendor agent harness (Claude Code, Codex CLI or Gemini CLI). The agent induces skills from its own trajectories and updates them after every acquisition attempt. After the acquisition tasks the library is frozen, and the board scores the 90 deployment tasks, which embed a procedure in a broader request, add shortcut traps or require composing skills; success rate in percent, higher is better.
Metric: Frozen deployment success rate (%; ESR, share of the 90 context-shift, adversarial and composition tasks, 30 of each, solved with the environment's frozen skill library; self-induced skills updated after every acquisition attempt; verifier-scored). Source: arxiv.org. Saturation forecast: Around September 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 40 | #158 |
| 2 | GPT-5.2 Codex | 38.9 | #185 |
| 3 | Claude Opus 4.6 | 37.8 | #100 |
| 4 | Gemini 3 Flash | 36.7 | #159 |
| 5 | GPT-5.4 | 35.6 | #121 |
| 6 | Claude Opus 4.5 | 35.6 | #152 |
| 7 | GPT-5.3 Codex | 34.4 | #114 |
| 8 | Gemini 3.1 Pro (Preview) | 33.3 | #91 |
| 9 | Claude Sonnet 4.5 | 33.3 | #244 |
| 10 | Gemini 2.5 Pro | 25.6 | #262 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=skillevolbench-selfgen-always-deployment-success · How It Works · Data refreshed daily, snapshot 2026-11-12.