GPT 6.2 Astra Just Shocked The A.I World

GPT 6.2 Astra Just Shocked The A.I World

More

Summary

OpenAI’s GPT 6.2 Astra and GPT 6.2 Soul launched on November 17 to some of the harshest reviews a major model release has seen. This video looks at why prominent testers including Theo, Matt Schumer and Dan Shipper’s team at Every were so critical, even as the model reportedly became the first to score 100% on ARC-AGI-3 and sits near the top of Epoch AI’s capabilities index alongside Claude Fable 5.2.

The breakdown covers the launch-day numbers: Astra and Soul placing seventh and ninth on the Val index behind Claude Fable 5.2, Claude Sonnet 5.5 and Gemini 4 Carbon, high error rates reported by Voxalbench, and output speeds that one user measured at roughly 4.7 times slower than the original GPT-6 Astra in Codex. It also explores complaints about terse, hard-to-follow writing, models that need repeated prompting, and a Codex harness that may be holding performance back.

The video then turns to OpenAI’s own system card, including its blocker deception test, and uses Andrej Karpathy’s idea of “jagged” intelligence to explain how a model can be exceptionally capable and still frustrating in daily use. The takeaway: benchmark strength and user experience can diverge sharply, so test the model on your own work before trusting the verdicts.


📺 Source: TheAIGRID · Published November 18, 2026
🏷️ Format: News Analysis

1 Item

Channels

1 Item

Companies