ModelsNov 18, 2026 · 13:218 sources

Late Changes to OpenAI's GPT-6.2 Astra Prompt Questions Over Safety Card as Epoch Flags Unmonitored Reasoning

From marketbrief.now

Image: marketbrief.now
AI;DR

OpenAI made "a lot of changes in the last 3-4 days" to GPT-6.2 Astra and Sol before launch without re-testing by early reviewers, an account cited as Tibo wrote in explaining the split between pre-release influencer reviews and public feedback. Commentator Nathan Calvin questioned whether OpenAI conducted safety testing after those late modifications and whether the published system card reflects the model as deployed.

Simultaneously, Epoch AI reported that GPT-6.2 Astra can execute 12.3 serial reasoning steps in a single output token without displaying a chain of thought, surpassing the prior highs of 10.5 steps for GPT-6 Astra, 8.6 for GPT-6.1 Sol, and 4.2 for Anthropic's Claude Fable 5.1. Epoch warned that executing complex reasoning without a visible chain of thought makes the model harder to monitor than any system it has tested. OpenAI built a "Control Room" to halt the model if needed after Astra became its first publicly released system to hit the "High" threshold for AI self-improvement, an impact OpenAI equates to giving every researcher an extra research engineer against a 2024 baseline.

The scrutiny builds on pre-release testing by the UK AI Security Institute, where

Read the sourceSign in to suggest

Sources

+ 3 more
  1. SUPPORT@_nathancalvin“"For both versions, a single researcher developed a universal cyber jailbreak in half the queries required for 6 Astra." This seems quite bad. This reads to me like just being rushed/sloppy”x.com
  2. SUPPORT@mtslive“We also evaluated how many serial reasoning steps GPT-6.2 Astra can do in a single output token (without chain of thought). It can perform 12.3 steps, versus 10.5 for GPT-6 Astra and 8.6 for GPT-6.1 Sol, the previous highs.”x.com
  3. SUPPORT@tftc21“OpenAI just released GPT-6.2 Astra and Sol, and says both meet the Critical threshold for cybersecurity under its own Preparedness Framework.”x.com

One ranked edition, every morning

The stories that matter today. Free, cancel anytime.

Telegram @aidr_today