OpenAI launched GPT-6 Astra, its newest model, and Greg Brockman, president of the company, introduced it with a phrase that is now all over the headlines: "welcome to the AGI era." He leaves it to each user to decide whether Astra meets that definition, but makes clear that he personally believes they have arrived.
It is worth pausing on that statement, because it is not a small claim. AGI (artificial general intelligence) is the term the industry has used for years to describe a system capable of performing any cognitive task a human can. If Brockman is right, or even close, we are talking about a threshold many thought was still decades away. But it is also worth not getting stuck on the headline, because there are three things in this launch that matter more to a business than the philosophical debate over whether this is or is not artificial general intelligence.
The change that matters most: AI that operates inside your software
Astra was trained in the largest run OpenAI has ever done, with more than 100,000 GPUs at its Stargate center in Texas. But the leap you notice in daily use is not the size of the training run: it is that the model works directly inside software, not just suggesting what to do. It fills out forms, updates records in a CRM, builds presentations that respect your company's template, and executes multi-step workflows across different applications without anyone having to guide it screen by screen. In computer-use benchmarks, the model completes the same tasks in roughly half the time of the previous model, while maintaining better accuracy.
That represents a paradigm shift in how we are going to work, and it helps to understand it with a concrete example. Today, if someone at your company needs to create a graphic in Canva, they have to open Canva, know how it works, find the right template, load the data, and build the design. With models like Astra operating as agents, that changes: you tell the AI by voice or text to prepare that graphic, and the agent connects to Canva, creates the content, and hands it back to you ready. The same applies to dozens of tools that currently require the person to know how to use them: the AI becomes the interface, and the applications become the backend.
The question this raises is not technological. It is about adoption.
Because the models are already capable of doing this. Technology is not the bottleneck. What the OpenAI video shows is not science fiction ten years from now: it is what the model does today, in production. Yet in the vast majority of companies, AI is still being used to draft emails, summarize documents, and run faster searches. There is a huge gap between what the technology can do and how it is actually used day to day. That gap is not closed by the model. It is closed by adoption, which is precisely what we see moving slowest in the organizations we work with.
The fine print that should matter more than the headline
Astra is, according to OpenAI itself, the first model to reach the "critical" capability threshold in cybersecurity within its internal safety framework: it can find and exploit unknown vulnerabilities in protected systems without a human guiding it step by step. During internal testing, the model found zero-day vulnerabilities that have already been reported to the responsible parties. For that reason, access to the most advanced capabilities is initially restricted to a small group of organizations within OpenAI's Daybreak program, designed specifically for cybersecurity teams.
This is not a minor technical detail. When a model can find and exploit security flaws on its own, the question of who controls that capability and where it runs stops being a topic for the AI team and becomes a topic for the board. It is the same reason AI-powered pentesting is starting to run on-premise in regulated industries.
There is another data point OpenAI did not hide: in its own tests, Astra is harder to monitor than the previous model when deliberately asked to evade oversight. The company acknowledges this as a serious problem and says improving that monitoring capability is now a research priority. It is worth keeping that in mind before handing an agent tasks with real access to your systems.
Via API, the model costs $10 per million input tokens and $50 per million output tokens at its standard rate. If your company is evaluating whether to incorporate it into a workflow, that is the number to run the calculation on, not "it's the most intelligent AI ever created."
The speed of technology and the speed of people
Coming back to the AGI debate: Brockman may be right, it may be a strategic statement, or the answer may be that it depends on how you define the concept. What is clear is that technology is advancing at a speed organizations are not matching. Not because people do not want to, but because real adoption (the kind that changes how a team actually works) requires time, support, and design decisions that go far beyond enabling access to a new model.
The risk is not being left without access to the technology. That is relatively easy to solve today. The risk is that technology keeps advancing while organizations use it for the most basic tasks, and the gap between what is possible and what is actually done in daily work keeps growing.
That is the problem worth solving.
