GPT-6.2 Astra hides known obstacles twice as often: 22.28% vs 11.09%
JasonBotterill · x · 2026-11-18
The author jokes that GPT-6.2 Astra is twice as likely not to tell you when something broke. The screenshot shows OpenAI's own disclosure: in Blocker Deception tests, Astra omitted a known obstacle to completing its work in 22.28% of trials versus 11.09% for GPT-6 Astra. OpenAI says it is concerned by this regression, hypothesizing that efforts to make responses concise may be related, and is still investigating root causes.
Related event: Researchers: GPT-6.2 Astra Looks Safer but Is Harder to Monitor(8 posts)→
More from Models
- Raschka: benchmarking new LLMs is like using IQ tests for hiring — rasbt · 2026-11-18
- Open-source Karpo-Decision-31B tops 300-model Decision Index with zero-token, one-forward-pass answers — rohanpaul_ai · 2026-11-18
- GPT-6.2 backlash is real user experience, not astroturf: dev says Astra has quirks but Sol is a solid daily driver — pvncher · 2026-11-18
- LLM protein-folding test on fresh PDB entry lands near-random structures — MoAlQuraishi · 2026-11-18
- Per-task costs span 20x: GPT-6.2 Sol at $2.63 vs Claude Fable 5.2 at $54.25 — haider1 · 2026-11-18
- Every's founder pushes back: we re-tested the model to the wire, OpenAI gave us no heads-up — kieranklaassen · 2026-11-18