Gemini 4 Argon finds 57/100 PR vulnerabilities for $166, second only to Grok 4.7
RexDouglass · x · 2026-10-14
BugBunny.ai's VulnPR-100 benchmark gave models 100 pull requests with known vulnerabilities, one hour each.
- Gemini 4 Argon found 57 for $166.40, ranking second overall; leader Grok 4.7 found 62 but cost $338.67, more than twice as much
- Argon beat GPT-5.6 Sol by 7 findings while costing $114 less
- Highlight: Argon caught a vulnerability every previously tested model missed, pushing combined benchmark coverage from 89 to 90/100
- New recall isn't cheap: mimo-v2.6-pro found 46 for $8.27 and DeepSeek V4.1 Flash found 44 for $12.79; Argon's extra 11-13 findings cost 13-20x more per review
Second place plus genuinely new coverage is still a strong result for Gemini 4 Argon.
More from Models
- Argon vs Astra vs Opus 5.5: flagship models battle on a squishy WebGPU physics demo — dreamwieber · 2026-10-14
- Redditor notes Chinese model releases slowed after distillation protections — Electrical_Rub_6009 · 2026-10-14
- Egocentric video bakeoff: Astra 59.4 beats Argon 55.3 and Opus 48.7 on action segmentation — RexDouglass · 2026-10-14
- Gemini 4 Argon beats Pokémon using less than half the tokens Gemini 3.8 Flash spent on 3 badges — waylaidwanderer · 2026-10-14
- Musk endorses Grok Bot, quoting a list of ways AI 'employees' beat humans — elonmusk · 2026-10-14
- Is Meta back at the AI frontier? 'Muse Wonder' rumor sparks Reddit buzz — RelevantEmergency707 · 2026-10-14