LibEvoBench - Evolving APIs: leaderboard
Metric: Average task performance (0-100) on evolving APIs (signature changed, introduced or deprecated within the covered versions): mean of four task metrics (API calling with version constraint, exact match; API calling parameter recall given name and version; API identification from a redacted docstring, exact match; signature recall F1) averaged over PyTorch, NumPy and SciPy versions, temperature 0 where supported and no thinking budget; higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 13 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5.4 | 86.4 |
| 2 | GPT-5.5 | 86.1 |
| 3 | Claude Sonnet 4.6 | 82.4 |
| 4 | Claude Sonnet 4 | 78.5 |
| 5 | Gemini 3 Flash (Minimal) | 78.5 |
| 6 | GPT-4.1 | 71.7 |
| 7 | GPT-5.1 | 71.6 |
| 8 | GPT-5 | 71.5 |
| 9 | Gemini 2.5 Flash (Non-reasoning) | 69.8 |
| 10 | Qwen 3.5 397B A17B (Non-reasoning) | 68.9 |
| 11 | Gemini 2.0 Flash | 67.5 |
| 12 | Qwen 3.5 122B A10B (Non-reasoning) | 62.3 |
| 13 | Qwen 3.5 35B A3B (Non-reasoning) | 54.7 |
Interactive version: theaggregate.ai/benchmark?slug=libevobench-evolving-apis · How It Works · Data refreshed daily, snapshot 2026-09-29.