Tool-Genesis (Direct) - Unit Tests (Hard): leaderboard
Metric: Unit-test score (0-1, times 100) over the standard plus boundary and negative tests, each test averaging a key-path F1 and an output embedding similarity, single-pass generation (Direct), Tool-Genesis's tool-creation tasks over 86 executable MCP servers (508 tools, 9,441 unit tests): from an abstract requirement the model writes the tool interface schema and an executable MCP server implementation; temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around 2028. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 | 16.1 | #131 |
| 2 | GPT-4.1 | 12.9 | #240 |
| 3 | DeepSeek V3.2 | 12.9 | #198 |
| 4 | Qwen 3 235B A22B 2507 Instruct | 10.8 | #291 |
| 5 | GPT-4.1 Mini | 8.8 | #346 |
| 6 | Kimi K2 | 8.7 | #236 |
| 7 | Qwen 3 32B | 7.8 | #424 |
| 8 | Qwen 3 14B | 7.5 | #524 |
| 9 | GPT-4o | 4.9 | #333 |
| 10 | Gemini 3 Flash (Preview) | 3.7 | #78 |
| 11 | Qwen 3 30B A3B 2507 Instruct | 2.5 | #464 |
| 12 | Qwen 3 8B | 0.1 | #667 |
| 13 | Qwen 3 4B | 0 | #823 |
| 14 | Claude 3.5 Haiku | 0 | #553 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=tool-genesis-direct-unit-tests-hard · How It Works · Data refreshed daily, snapshot 2026-10-11.