C3-Bench (Controllable Code Completion) - Implementation Control: leaderboard
C3-Bench (Alibaba Group with USTC, UCAS and CASIA, 2026) gives a model the Python code before and after a missing middle segment together with an instruction fixing how that segment must be implemented (a required algorithm, data structure, control flow or parameter handling), over 1,286 Implementation-Control tasks built from HumanEval and SAFIM. A completion counts only when it passes the unit tests and a Claude3.5-Sonnet judge finds the instructed implementation; the score is the percentage of all tasks that count, higher is better. Base code models are prompted in fill-in-the-middle token format and chat models in a ChatML template, and the authors' Qwen2.5-Coder-C3 models are fine-tuned on synthetic data in this task format.
Metric: Instruction-Following Rate (%; 1,286 ICC tasks: passes the unit tests and a Claude3.5-Sonnet judge finds the instructed implementation). Source: arxiv.org. Saturation forecast: Estimated already saturated. 42 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 (2024-12-17) | 62.9 | #310 |
| 2 | Claude 3.5 Sonnet (20241022) | 60.9 | #635 |
| 3 | GPT-4o (2024-11-20) | 59.6 | #864 |
| 4 | Gemini 2.0 Flash | 59.5 | #702 |
| 5 | GPT-4o (2024-08-06) | 58.6 | #731 |
| 6 | O1 Preview | 57.7 | #483 |
| 7 | O1 Mini | 55 | #761 |
| 8 | DeepSeek V3 (0324) | 53 | #704 |
| 9 | DeepSeek V3 | 47.3 | #705 |
| 10 | GPT-4o Mini (2024-07-18) | 42.2 | #1512 |
| 11 | GPT-4o (2024-05-13) | 41.3 | #734 |
| 12 | Claude 3.5 Haiku | 40.7 | #1563 |
| 13 | Qwen 2.5 Coder 32B | 38.7 | #1869 |
| 14 | Qwen 2.5 Coder 14B | 36.2 | #3340 |
| 15 | Codestral-22B-v0.1 | 34.1 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-12.