Summary
OpenAI’s release of its two new flagship models, GPT-6.2 Astra and Soul, has been met with a wave of negative early reviews, followed by a counternarrative that the models are actually strong. Matthew Berman, who was part of OpenAI’s early access testing program, walks through what happened and gives his own verdict after spending hundreds of billions of tokens testing the models.
His assessment is that Astra is a fine model that excels mainly at computer use, where he says OpenAI remains dominant. He demonstrates a 3D replica of Tokyo built in Unreal Engine with tens of thousands of hand-built buildings, noting how the model needed repeated nudging along the way. He also shares a case in which Astra used computer control to diagnose a long-running frame-drop problem by reading OBS logs and forcing a buffer setting on, fixing an issue that had persisted for 18 months.
The video covers the controversy itself, including reactions from early-access testers such as Matt Shumer and Theo and the claims of reviewer collusion. Berman concludes that, compared with Anthropic’s Opus and Fable, the new OpenAI models fall short, and that Anthropic’s subscription, with generous Opus 5.5 quota, is now the better value.
📺 Source: Matthew Berman · Published November 19, 2026
🏷️ Format: Review







