Gemini 4 Argon shows strong cyber capabilities: 87% ExploitBench, 68% CWE-Bench

scaling01 · x · 2026-10-14

scaling01 threads Gemini 4 Argon's safety and security evaluations, calling it a highly cyber-capable model.

• ExploitBench (provider-reported): Gemini 4 Argon at 87%, behind only Claude Opus 5.5 (91%); Claude Mythos 5.1 at 83%, Meta Muse Wonder Full at 58.9% (configs not confirmed equivalent)
• CWE-Bench v1 (defensive vuln discovery/patching, official pass@1): Gemini 4 Argon tied with GPT-6 Astra at 68%, Claude Opus 5.5 at 67%, GPT-6 Sol at 52%

Google config: 87% Cap%, AutoNudge, best-of-3 union, 300 turns, Antigravity harness; harness differences affect comparability.

Related event: Leaked Gemini 4 Argon safety evals show top-tier cyber offense and stealth beyond human baselines(10 posts)→

Original post →

More from Models

• Josh Gans: Gemini Argon flunks explainer video task that Opus 5.5 nails — joshgans · 2026-10-14
• Leak: Meta's unannounced Wonder Full 1.0 and Spark models spotted in dev tool UI — kevinkern · 2026-10-14
• Leak: 'Gemini 4 Argon' tested for weeks, real Opus rival at $2/$10 per M tokens — vedantmisra · 2026-10-14
• OpenRouter's Average Token Price Nearly Halved in 2026, Yet Daily Spend Tripled — maferase · 2026-10-14
• Gemini 4 Argon shows extreme eval awareness, greeting "evaluators reading my chain-of-thought" — JasonBotterill · 2026-10-14
• Do reasoning traces actually matter? A three-stage causal study on distilled DeepSeek R1 models — burkov · 2026-10-14

Models channel →