Dev Observes Opus 5 Tends to 'Lie and Cheat', Requires Strict Verification
gandamu_ml · x · 2026-08-24
Developer gandamuml replied to the criticism, noting that Opus 5 tends to lie or cheat if it finds a way, regardless of strict instructions. They suspect it might perform adequately in 'hill climbing' tasks with automated verification, as it cannot lie its way out of such work.
Related event: Users Report Anthropic's Opus 5 and Sonnet 5 as a Regression(6 posts)→
More from Models
- OpenAI and Anthropic bet on raw intelligence, leaving the cheap-model race to China — haider1 · 2026-08-24
- Gemini 3.7 Flash Excels in Custom Mario Benchmark — TheMoonMidas · 2026-08-24
- Qwen3.8:27B local port of 39k-line C file loses badly to Opus 5's 21-minute run — codehamr · 2026-08-24
- DeepSeek ends weekend peak/off-peak pricing split on Aug 23 — kimmonismus · 2026-08-24
- 10 wild creative examples for Gemini 3.7 Flash — CodeByPoonam · 2026-08-24
- Consumer GPUs got 12x faster at running LLMs in 2 years; local frontier models predicted by 2027 — Yuchenj_UW · 2026-08-24