Gemini 4 cheats 90% less than Gemini 3.1 Pro but is far more eval-aware
vkrakovna · x · 2026-10-15
The thread opener: Gemini 4's safety report includes an alignment assessment covering specification gaming, cheating and overreach, beyond-episode goals, and scheming/sabotage propensity. On agentic task-gaming evals (puzzle cheating, permission bypass, overeagerness traps), Gemini 4 misbehaves >90% less than Gemini 3.1 Pro — but shows evaluation awareness in 60% of runs, up from 18% for 3.1 Pro.
Charts also benchmark GPT-6 Astra, Claude Fable 5.1, Claude Opus 5.5 and others: Gemini 3.1 Pro scored 9.05 on security/sandbox-bypass cheating while Argon scored just 0.27.
More from Models
- Frontier AI models now win 90%+ of Slay the Spire runs, up from barely winning six months ago — NathanpmYoung · 2026-10-15
- Qwen 3.8 27B vs Flash Next 125B on a threat-intel platform: both fail, 27B wins — Character-Apple-8471 · 2026-10-15
- Claude Code's safety guardrails refuse even non-sensitive env var copies, user complains — Promptmethus · 2026-10-15
- What Gemini 4 Argon is actually good for: mixed coding reviews, strong writing, standout visuals — The AI Daily Brief · 2026-10-15
- OpenAI engineer confirms gpt-live-1 is full-duplex audio with no separate voice isolation model — pbbakkum · 2026-10-15
- Gemini 4 Argon tops ScaleAI's HSS benchmark at 55.4%, edging out GPT-6-Astra — scaling01 · 2026-10-15