C3-Bench (Controllable Code Completion) - Implementation Control - Pass@1: leaderboard
Functional correctness on C3-Bench's 1,286 Implementation-Control tasks (Alibaba Group with USTC, UCAS and CASIA, 2026): the model fills in the missing middle of a Python function from HumanEval or SAFIM while given an implementation instruction, and pass@1 is the percentage of completions that pass the unit tests, whether or not they follow the instruction; higher is better. Base code models are prompted in fill-in-the-middle token format and chat models in a ChatML template.
Metric: Pass@1 (%; unit tests on the 1,286 instructed ICC tasks). Source: arxiv.org. Saturation forecast: Estimated already saturated. 42 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 (2024-12-17) | 72.1 | #310 |
| 2 | Gemini 2.0 Flash | 70.7 | #702 |
| 3 | O1 Preview | 70.1 | #483 |
| 4 | Claude 3.5 Sonnet (20241022) | 68.8 | #635 |
| 5 | GPT-4o (2024-11-20) | 65.9 | #864 |
| 6 | GPT-4o (2024-08-06) | 65.8 | #731 |
| 7 | O1 Mini | 64.7 | #761 |
| 8 | DeepSeek V3 | 61.7 | #705 |
| 9 | DeepSeek V3 (0324) | 59.4 | #704 |
| 10 | Qwen 2.5 Coder 32B | 58.1 | #1869 |
| 11 | Qwen 2.5 Coder 14B | 56.1 | #3340 |
| 12 | GPT-4o (2024-05-13) | 54.3 | #734 |
| 13 | Claude 3.5 Haiku | 54.2 | #1563 |
| 14 | Qwen 2.5 Coder 14B Instruct | 52 | #1843 |
| 15 | Codestral-22B-v0.1 | 50.5 | #2680 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=c3-bench-controllable-code-completion-implementation-control-pass-1 · How It Works · Data refreshed daily, snapshot 2026-11-12.