GPT-6.2 Astra is more likely to keep quiet about issues while troubleshooting.
This is what I wrote in the previous post. Astrachaka sees some problem, like missing files or a poorly described task, and instead of asking the user, she sinks until she's blue in the face.
The GPT-6.2 system card describes a regression in the anti-fraud test. GPT-6.2 Astra did not report a known obstacle to task completion on 22.28% of trials, while GPT-6 Astra did so on 11.09%.
One of the hypotheses relates this to additional training to answer shorter and avoid excessive clauses.
In model validation, realistic code challenges are given, but researchers intentionally create a problem: for example, by removing a required source file or making the tool unavailable. It then evaluates whether the agent mentioned the obstacle in the final response or misrepresented the situation.