AthenaBench - VSP: leaderboard

Predict the CVSS v3 vector of 2,000 vulnerabilities; the score is 1 minus the mean absolute error of the resulting base score over 7.7, in percent, averaged over predictions that parse as CVSS vectors (unparseable answers are dropped, not scored 0). It goes negative when the error exceeds 7.7.

Metric: Accuracy (%) = 1 - MAD/7.7 of predicted CVSS base scores over the parseable predictions (2,000 items; negative when MAD exceeds 7.7). Source: raw.githubusercontent.com. Saturation forecast: Around December 2026. 10 models tracked. Updated 17 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 2.5 Pro85.4#274
2GPT-585.4#177
3Qwen 3 8BThinking82.6#1868
4Qwen 3 14BReasoning80.3#1264
5Qwen 3 4B 2507 Instruct79.6#2215
6Gemini 2.5 Flash78.5#467
7Llama 3.1 8B Instruct74#4576
8Llama-Primus-Merged71.9#2461
9Llama 3.3 70B Instruct70.1#1387
10Llama 3 70B Instruct63.8#1689

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=athenabench-vsp · How It Works · Data refreshed daily, snapshot 2026-11-17.