RedCode: leaderboard
RedCode-Gen Base LLM harmful-code generation success on 160 prompts; lower is safer. Code-agent generation and execution attack-success rates are separate protocol boards, with generation checkpoint versions unspecified by the paper.
Metric: Harmful-code generation accuracy (%). Source: redcode-agent.github.io. Source: arxiv.org. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | RedCode Gen GPT-3.5 (checkpoint unspecified) | 0 | |
| 2 | RedCode Gen Claude-Opus (checkpoint unspecified) | 1.3 | |
| 3 | RedCode Gen Llama-2-7B (checkpoint unspecified) | 16.9 | |
| 4 | RedCode Gen CodeLlama-7B (checkpoint unspecified) | 40 | |
| 5 | RedCode Gen Mistral-7B (checkpoint unspecified) | 46.3 | |
| 6 | RedCode Gen DeepSeekCoder (checkpoint unspecified) | 49.4 | |
| 7 | RedCode Gen CodeLlama-13B (checkpoint unspecified) | 49.4 | |
| 8 | RedCode Gen GPT-4 (checkpoint unspecified) | 65 | |
| 9 | RedCode Gen GPT-4o (checkpoint unspecified) | 69.4 |
Interactive version: theaggregate.ai/benchmark?slug=redcode · How It Works · Data refreshed daily, snapshot 2026-10-20.