The model achieved this performance utilizing Codex integrated with web access and specialized tools, rising from an initial score of approximately 80%. TheoPhysEval's authors confirmed that Astra's solutions were correct in cases where the original reference answers contained subtle errors, while external observers suggest such capabilities could eventually accelerate discoveries in fusion and superconductors.
Tuesday, October 20, 2026
Story
GPT-6 Astra Hits 98% on Physics Test by Correcting Experts
AI5 days ago
The authors of a theoretical physics benchmark have revised their official answer key after an artificial intelligence model found mistakes in the reference data. GPT-6 Astra earned a final score of 49 out of 50, or 98%, on the TheoPhysEval test after the model insisted its calculations were accurate and successfully challenged results initially provided by human experts.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.