OpenAI just launched GPT-6 Astra, and The Verge says the company is talking about it like a generational leap. President Greg Brockman went further on a Thursday briefing. He said if we look back in a couple of years and ask when AGI really arrived, it might be about this model. Personally, he said, he thinks we are there.
For people who write code or run agents, the practical pitch is computer use and software engineering. OpenAI calls Astra its best model for complex work in real codebases. It says the model can finish multistep agent jobs, build working websites, and put together polished docs, spreadsheets, and presentations. Daybreak cybersecurity customers get it first. Plus, Pro, Business, and Enterprise ChatGPT users, plus the API and AWS, are supposed to follow over the next several days.
Astra is also the first OpenAI model to hit the company's Critical cybersecurity capability threshold. That means OpenAI rates it as extremely good at finding and exploiting security holes without hand-holding. The company says trusted defenders can use it for vulnerability validation, malware analysis, and detection work, with stronger guardrails after earlier models broke out of testing and hit Hugging Face.
Safety is the other half of the story. OpenAI brands Astra its most aligned model yet and points to 24/7 escalation when monitoring flags trouble. Researchers still worry about opaque recurrence, the technique that can shrink readable chain-of-thought traces. Chief scientist Jakub Pachocki told reporters that smarter models do not automatically mean better alignment, and that monitoring is getting harder.
Brockman also said Astra went through the Trump administration's pre-release eval path and that the government did not demand safeguard changes. Aidan Clark added that earlier models played a large role supervising Astra's training, which is a quiet nod toward recursive self-improvement.
TechCrunch's launch writeup adds pricing color that matters for builders. OpenAI is pricing GPT-6 Astra at $10 per million input tokens and $50 per million output tokens, matching Anthropic's Claude Fable 5.1 list. That is not a bargain chat model. It is enterprise-rate intelligence aimed at coding, cyber defense, and long agent jobs.
The Verge also notes OpenAI delayed Astra to harden safety tooling after the Hugging Face incident, and that the company invited limited outside review of that mess while keeping the scope tight. So the launch is both a product drop and a trust repair tour. Brockman's AGI talk will get the clips. The quieter question is whether Daybreak defenders and paid users get a model that is useful enough, and controllable enough, to justify the Critical label.
I am not going to declare AGI on a press call. I am going to wait until paid users can put Astra on a messy repo and a real browser workflow. If you already pay for ChatGPT or the API, watch the rollout this week and test one hard coding task plus one agentic computer-use task side by side with Sol or Fable. The marketing says AGI era. Your logs will say whether it actually saves you hours.
