Downstream task performance on SkillsBench
51.1Pass RateGold Skills
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Gold SkillsBackbone Model=GPT-5.2-Codex2026.06 | 51.1 | 1.12 | |
| Gold SkillsBackbone Model=Gemini-3-Pro2026.06 | 48.4 | 1.18 | |
| SkillComposerBackbone Model=GPT-5.2-Codex2026.06 | 45.3 | 1.03 | |
| Retrieval (oracle)Backbone Model=GPT-5.2-Codex2026.06 | 44 | 1.13 | |
| Retrieval (top-3)Backbone Model=GPT-5.2-Codex2026.06 | 44 | 1.09 | |
| SkillComposerBackbone Model=Gemini-3-Pro2026.06 | 44 | 1.08 | |
| Retrieval (oracle)Backbone Model=Gemini-3-Pro2026.06 | 42.2 | 1.19 | |
| Retrieval (top-3)Backbone Model=Gemini-3-Pro2026.06 | 41.8 | 1.14 | |
| All SkillsBackbone Model=Gemini-3-Pro2026.06 | 38.7 | 1.33 | |
| All SkillsBackbone Model=GPT-5.2-Codex2026.06 | 29.3 | 1.27 | |
| No SkillsBackbone Model=Gemini-3-Pro2026.06 | 25.8 | 0.99 | |
| No SkillsBackbone Model=GPT-5.2-Codex2026.06 | 22.2 | 0.94 |