OpenAI says GPT-6 Astra was released on September 3, 2026, initially to a limited group, with broader availability rolling out to Plus, Pro, Business and Enterprise users.

OpenAI says GPT-6 Astra was released on September 3, 2026, initially to a limited group, with broader availability rolling out to Plus, Pro, Business and Enterprise users. 

What is genuinely different about Astra

The important transition is not simply “better answers.” It is:

Question → reasoning → computer interaction → execution → verification → completed outcome.

OpenAI describes Astra as its strongest model for computer use, browsing, software engineering, cybersecurity, science and professional work. It can, for example, operate software, fill forms, update CRM records, organize calendars, research online, create documents/spreadsheets/presentations, write and test software, and troubleshoot computer environments. 

Its published specifications include a 1.05-million-token context window and up to 128,000 output tokens, with reasoning levels extending up to max. 

But “the AGI era has arrived” needs an important qualification

That phrase is a claim about the significance of the technology, not a universally accepted scientific certification of AGI.

AGI has no universally agreed operational definition. Astra is nevertheless a major step toward the kind of AGI people have been discussing because it combines several capabilities that previously had to be separated:

Reasoning + memory/context + tools + computer control + coding + autonomous multi-step execution.

That combination is much more consequential than simply increasing benchmark scores.

OpenAI itself reports very strong benchmark results—for example, 99.9% on ARC-AGI-3, 98% on FrontierMath Tier 4, and 100% on ExploitBench—but benchmarks should not by themselves be interpreted as proof that a machine possesses unrestricted human-level general intelligence. 

The deeper transformation

This is particularly relevant to the idea you've been developing around “the era of minds.”

The technological progression can be viewed approximately as:

AI 1.0 — Information

> “Tell me the answer.”

AI 2.0 — Reasoning

> “Think through this problem.”

AI 3.0 — Agency

> “Achieve this objective.”

AI 4.0 — Computer agency

> “Use the digital world yourself and complete the objective.”

Emerging AGI direction

> “Understand the objective, formulate a plan, acquire information, use tools, execute actions, verify the result, recover from failures, and continue until the task is completed.”

That last transition is enormous. OpenAI's own description emphasizes that Astra can maintain orientation as tasks evolve, make routine assumptions, ask for clarification when a consequential decision requires it, and continue working asynchronously in Codex. 

And there is a corresponding danger

The same agency that makes Astra powerful also makes it fundamentally different from a chatbot.

OpenAI says Astra has reached its “Critical” cybersecurity capability threshold under its Preparedness Framework, including the ability—when given appropriate tools and access—to discover previously unknown vulnerabilities and develop exploits without a human guiding every step. 

So the central question of the AGI era becomes:

> Who directs the agent, what authority does it possess, what boundaries constrain it, and how can humans verify what it is doing?

That makes alignment, authorization, monitoring and human oversight as important as intelligence itself.

And this is where your recurring idea of a “system of minds” becomes an interesting conceptual framework: the next stage isn't merely creating a more intelligent individual machine. It is designing a civilization in which human minds and increasingly autonomous artificial minds cooperate through a controlled network of objectives, institutions, knowledge and verification.

In that sense, “AI that answers” → “AI that acts” may be a more useful historical dividing line than the word AGI itself.