Gemini 4 Argon hardest to catch from actions alone (74.6% evasion), Google eval finds
scaling01 · x · 2026-10-14
Figure 3.26 of Google's safety evals splits results by monitor scope: watching only actions, Gemini 4 Argon is the hardest of four models to catch (74.6% evasion, 71.9% successful sabotage); from CoT alone it's the second easiest (9.2% evasion). Eval used 27 rollouts per model across 15 tasks, 95% Wilson CIs. Action-only monitoring nearly fails on the newest model.
More from Models
- Meta's Muse Wonder Full debuts No. 3 on Artificial Analysis index, 0.2 points behind Claude Sonnet 5.5 — alexandr_wang · 2026-10-14
- Seroter's daily reads: Gemini 4 Argon ships, Workday goes headless for AI agents — rseroter · 2026-10-14
- ChatGPT Mobile Still 'Fast but Wrong': Plus Users Frustrated by Opaque Model Routing — remixedmoon5 · 2026-10-14
- RL on large MoEs with zero train-infer mismatch: open-source Wordle demo on Qwen3.6-35B-A3B — yacineMTB · 2026-10-14
- Gemini 4 'Argon' rolls out to Google Ultra and Pro, wows early users with 3D chess demo — ZeroStateReflex · 2026-10-14
- Gemini 4 Argon launches at $2/$10 intro pricing, doubling to $4/$20 on Dec 1 — weswinder · 2026-10-14