GPT-6 Astra's benchmark results on ARC-AGI-3, math and OSWorld, collected
A video post lists Astra's ARC-AGI-3, open math problem, zero-day vulnerability and 72.6% OSWorld results, and says developers still cannot fully understand its reasoning.
You continue to calmly scroll through the feed and argue about whether the era of general artificial intelligence has already arrived. Meanwhile, the new OpenAI GPT-6 Astra silently beats out 99.9% on ARC-AGI-3. It solves complex open mathematical problems. It finds previously unknown zero-day vulnerabilities. And it reaches 72.6% in computer control via OSWorld. The line between a regular chatbot and a full-fledged autonomous operator is rapidly disappearing. The combination of unprecedented benchmark results, the ability to independently perform multi-stage tasks in the cloud and real scientific discoveries proves one thing. We are standing on the verge of a historic shift. Artificial intelligence is no longer just a text generator. It is turning into an independent force capable of performing real work. Work that humans will increasingly delegate to autonomous systems. At the same time, developers themselves are already sounding the alarm. The reason is simple: under difficult challenges, they still cannot fully control or understand its reasoning. This is no longer a theory about some abstract AGI from the distant future. It is the harsh reality of an engineering breakthrough happening right now. It will either permanently change the rules of the game for those who learn to use these new tools in time. Or it will sweep aside those who still perceive autonomous AI systems as nothing more than toys for writing posts.
This post quotes another post: x.com/kimidyor/status/2095628832925462913