🔮 Astra outruns visibility EV#600

Welcome to the 600th Sunday edition of Exponential View, where we've been analyzing AI for the past eleven years. As a special celebration, we're offering a limited-time discount: 60% off the first year of your annual membership. This is the biggest discount we'll provide and the lowest price you'll ever get for Exponential View, as we review our prices this fall. The offer is open for 24 hours only, so don't miss out on this fantastic deal.

OpenAI's GPT-6, Astra, has recently outperformed other leading models, such as Claude Fable 5.1, on various benchmarks. In fact, Astra has been working diligently on tidying up my Desktop and Download folders, with Fable 5.1 also demonstrating impressive speed in conducting useful analysis that once required multiple steps and occasional human intervention.

Astra's capabilities are particularly noteworthy when it comes to tackling complex mathematical problems. It completes tasks 30.9 minutes faster than GPT 5.6 Sol. Mathematician Bartosz Naskręcki has expressed his enthusiasm, stating that Astra's performance feels like the arrival of a new era where mathematicians can focus solely on ideation and exploration.

Real-world demonstrations reveal a significant performance jump on technical and design tasks. Astra has shown remarkable efficiency in developing simulated worlds inhabited by agents communicating and collaborating together, as well as creating 3D models of Zillow listings. On the ARC-AGI-3 benchmark, Astra requires fewer actions than the average human on 96% of the levels it completes, averaging 51.7% fewer actions per level. This performance defies the original researchers' expectations.