WHAT NEW IN X10.29 THU

AI collections from X · 22 collections · 60,636 posts

gpt-6POST

GPT-6 Astra's key shift is acting, not just thinking, thread says

A retold thread says Astra moves beyond text-in, text-out to browsing, clicking, filling forms and running multi-step workflows on its own, and lists scores such as FrontierMath and ARC-AGI-3.

Jack
@jackcoder0Archived 2026.09.14Originally posted 2026.09.05
Jack @jackcoder0Originally posted 2026.09.05
Open original on X

His friend started by explaining why GPT-6 Astra is fundamentally different from every ChatGPT model before it. "Every model from GPT-3.5 through GPT-5.6 operates on the same principle you type a prompt, the model generates a response, you read the response. Text in, text out. The model thinks. It doesn't act. It can tell you how to book a flight. It can't book the flight. It can write code. It can't run the code. It can explain how to update a spreadsheet. It can't open the spreadsheet and update it." He told him GPT-6 Astra breaks that boundary. OpenAI calls it their "most intelligent and aligned model" but intelligence isn't the headline. Agency is. Astra can use a computer the way a human uses a computer browsing websites, clicking buttons, reading screens, filling forms, navigating applications, and executing multi-step workflows autonomously. The model doesn't just think about the task. It performs the task. He said the shift from "answering" to "doing" is the most significant change in ChatGPT's history more significant than the jump from GPT-3.5 to GPT-4, more significant than the addition of vision, more significant than voice mode. Those changes made the chatbot smarter, more capable, and more natural. GPT-6 Astra makes it operational. The chatbot becomes an operator. He told him the benchmark numbers tell part of the story. Astra scores 98% on FrontierMath Tier 4 the hardest mathematical reasoning problems that took professional mathematicians weeks to solve. It scores 99.9% on ARC-AGI-3 the test designed to measure artificial general intelligence. It scores 100% on ExploitBench a cybersecurity benchmark no previous model has saturated. But the benchmark numbers describe intelligence. The features describe what that intelligence can now do in the real world. Here are the 11 capabilities he showed him.

x.com/jackcoder0/status/2096179206665248866Archived 2026.09.14 16:18