GPT-5.6 Sol High feels 40-50% less reliable since Astra rollout, business user reports

biarschal95 · reddit · 2026-09-08

A Reddit user building an AI-heavy recruiting SaaS reports GPT-5.6 Sol High/Very High has degraded sharply since the Astra rollout — not on coding or benchmarks, but on long-form business reasoning and strategic context retention, which he estimates lost 40-50% reliability. Symptoms: contradicting a summary it generated 10 seconds earlier, reverting to corrections it already conceded, and 5-10 minute reasoning loops on simple state fixes. He notes the model remembers individual facts but misconnects them and builds confidently on the mistake, and asks whether this is a nerf, rollout instability, or something others have seen after past releases.

Original post →

More from Models

Models channel →