DevBench - Acceptance Testing: leaderboard
Each model writes acceptance tests (test inputs and expected outputs) from the requirements, design and reference source code for 22 repositories, with one review round; the score is the share of those tests that pass when run against the reference implementation, averaged uniformly over repositories. It grades the tests a model writes, not the code.
Metric: Oracle Test (%). Source: raw.githubusercontent.com. Saturation forecast: Estimated already saturated. 8 models tracked. Updated 17 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 Preview (0125) | 29.2 | #1197 |
| 2 | GPT-4 Preview (1106) | 25.9 | #980 |
| 3 | DeepSeek-Coder-33B-Instruct | 13.6 | #2726 |
| 4 | GPT-3.5 Turbo (1106) | 11.7 | #3361 |
| 5 | CodeLlama-34B-Instruct | 4.5 | #7944 |
| 6 | deepseek-coder-1.3B-instruct | 0 | #9476 |
| 7 | CodeLlama-7B-Instruct | 0 | #9490 |
| 8 | CodeLlama-13B-Instruct | 0 | #8718 |
No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.
Interactive version: theaggregate.ai/benchmark?slug=devbench-acceptance-testing · How It Works · Data refreshed daily, snapshot 2026-11-17.