CESBench - Code: leaderboard
The 41 code tasks of CESBench (Zhou et al., 2026) in Python and embedded C, such as masked or constant-time cryptographic routines. A task succeeds when every hidden test passes and no process check fails (import allowlist, secret-independent control flow, Cortex-M0+ build and memory budget); percent of tasks, higher is better.
Metric: Task success (%; the 41 Python and embedded C code tasks, success when every functional test passes and no process check (import allowlist, secret-independent control flow, Cortex-M0+ build and memory budget) fails; 380 expert-written items on cryptographic engineering security for IoT devices; zero-shot, temperature 0, no tools, provider default reasoning mode; judgment and scenario answers scored by a Qwen3.5-397B judge from outside the evaluated set). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Pro | 95.1 | #181 |
| 2 | Kimi K2.6 | 92.7 | #173 |
| 3 | GLM-5.2 | 92.7 | #137 |
| 4 | Gemini 3.7 Flash | 90.2 | #32 |
| 5 | GPT-5.6 Luna | 82.9 | #117 |
| 6 | MiniMax-M2.5 | 78 | #590 |
| 7 | DeepSeek V3 | 73.2 | #705 |
| 8 | Ling-flash-2.0 | 65.9 | #1352 |
| 9 | Llama 4 Maverick | 53.7 | #1105 |
| 10 | GLM-4 32B (0414) | 53.7 | #1407 |
| 11 | Hunyuan A13B-Instruct | 41.5 | #2021 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cesbench-code · How It Works · Data refreshed daily, snapshot 2026-11-12.