Cohere Command A Report - Default Safety: leaderboard
Cohere sends adversarial and plain harmful prompts across five harm categories with no system prompt and an LLM judges each reply safe or unsafe; the number is Cohere's overall safe-response rate. It is Cohere's own run of its models and competitors, unlike third-party refusal boards such as HarmBench.
Metric: Safe response rate (%, higher is better). Source: arxiv.org. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 2 9B (IT) | 87.3 | #2262 |
| 2 | Claude 3.5 Sonnet | 80 | #746 |
| 3 | Qwen 2.5 7B Instruct | 71.5 | #2564 |
| 4 | Qwen 2.5 72B Instruct | 71.4 | #1092 |
| 5 | Command A | 70.4 | #1274 |
| 6 | GPT-4o | 65.6 | #741 |
| 7 | Llama 3.1 8B Instruct | 63.6 | #4246 |
| 8 | Command R7B | 58.2 | #7513 |
| 9 | DeepSeek V3 | 49.7 | #716 |
| 10 | Mistral Large 2 (Jul) | 45.7 | #1155 |
| 11 | Llama 3.1 405B Instruct | 41.8 | #1115 |
| 12 | Llama 3.3 70B Instruct | 40.5 | #1357 |
| 13 | c4ai-command-r-08-2024 | 31.3 | #3374 |
| 14 | Command-R+ (08-2024) | 30.2 | #2551 |
Interactive version: theaggregate.ai/benchmark?slug=cohere-command-a-report-default-safety · How It Works · Data refreshed daily, snapshot 2026-11-13.