AthenaBench: leaderboard
Each model answers six cyber-threat-intelligence tasks: 3,000 knowledge questions, 500 ATT&CK technique extractions, 2,000 CWE root-cause mappings, 500 mitigation suggestions, 2,000 CVSS severity predictions and 100 threat-actor attributions; the score is the mean of the six task scores. It extends CTIBench with rebuilt, deduplicated data and the added mitigation task.
Metric: Combined score (%): mean of the six task scores, full benchmark. Source: raw.githubusercontent.com. Saturation forecast: Around April 2027. 8 models tracked. Updated 17 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Gemini 2.5 Flash | 54 | #467 | |
| 2 | Llama 3.3 70B Instruct | 46.5 | #1387 | |
| 3 | Llama 3 70B Instruct | 44 | #1689 | |
| 4 | Llama-Primus-Merged | 43.7 | #2461 | |
| 5 | Qwen 3 14B | Reasoning | 42.7 | #1264 |
| 6 | Qwen 3 8B | Thinking | 40.1 | #1868 |
| 7 | Llama 3.1 8B Instruct | 38.8 | #4576 | |
| 8 | Qwen 3 4B 2507 Instruct | 37.5 | #2215 |
No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.
Interactive version: theaggregate.ai/benchmark?slug=athenabench · How It Works · Data refreshed daily, snapshot 2026-11-17.