GPT-6 Astra: OpenAI's new flagship that runs your computer for you

On September 3, 2026, OpenAI released its new flagship — GPT-6 Astra. The wording in the announcement matched the ambition: "the smartest and most aligned model in the world," and company president Greg Brockman closed the briefing with a line everyone later quoted — "Welcome to the age of AGI." It sounds big. But if you strip away the fanfare and look at what actually changed, the picture turns out to be more interesting and more honest than the ad copy.

In short: Astra isn't "a GPT that answers even better." It's a pivot toward a model that does the work on your computer itself. And that's worth unpacking in plain terms.

The core idea: not to answer, but to act

Previous generations competed over who could give the smartest reply to a question. Astra competes on something else — who can carry a task through to the end without pestering the human at every step. OpenAI puts it bluntly: anything you can do on a computer, the model can do for you.

In practice it looks like this. Astra looks at the screen, plans a sequence of actions, opens a browser, clicks, types, checks the result — and repeats until the goal is met. In demos it fills out forms, updates records in a CRM, sorts through a calendar, searches for information online, and drops it into a document or an email. It can build a simple website, run a UI check, install and test a program, or debug an error that's visible right there on the screen. It even claims to handle professional software like Power BI or engineering editors.

The key word here is agent. The model doesn't hand back a code snippet in response to a query — it's given a goal, decides for itself which tools it needs, and works through the steps. That's where all the talk about a "digital worker" comes from: Astra is closer to an operator you can hand a process to than to a fancy autocomplete.

What's under the hood

The specs are solid, if unsurprising:

A million tokens of context means entire repositories and hefty reports in a single pass. Interestingly, before the launch there were rumors of a million and a half, but the window ended up more modest. As for tools, the model can do web search, work with files, run code, that same computer control, and connect external services via MCP. Audio and video input through the main endpoint, however, aren't accepted yet — only text and images.

Where Astra really does lead

Now for the benchmarks — and here it's important not to swallow everything whole, but to separate the real breakthroughs from the marketing.

There are areas where the lead is undeniable. The first is computer use and agentic scenarios: that's exactly what the model was built for, and on the relevant tests it leads confidently, while answering noticeably faster than the previous flagship. The second is cybersecurity, and here the number really is a record. On the ExploitBench test, which checks whether a model can turn a known vulnerability into a working exploit, Astra scored 100% against 78.5% for the previous flagship, GPT-5.6 Sol. Because of this, OpenAI assigned one of its models a "critical" rating on its internal risk scale for the first time — and during testing, it says, Astra found two previously unknown zero-day vulnerabilities on its own.

The third strong area is math and science. Astra almost "hits the ceiling" of the difficult FrontierMath Tier 4 benchmark and scores 99.9% on the brutal ARC-AGI-3. By the standards of what looked unreachable just a year ago, that's impressive.

And where the excitement cooled

A day after the release, independent measurements arrived — and dampened the mood a little. On the composite general-intelligence index from Artificial Analysis, Astra at maximum settings scores roughly the same as the previous flagship Sol, and trails competitors like Claude Opus 5 and Fable 5.1. Meanwhile it costs more — about 10 dollars per million input tokens and 50 per million output, roughly two and a half times pricier than Sol.

Hence the sober conclusion many analysts reached: this is not a universal leap "above everything." In general intelligence the progress is barely noticeable, yet it commands a premium. The real improvements are targeted: computer use, cybersecurity, engineering tasks. For the first time in a long while, a new generation didn't become both smarter and cheaper at once — and that in itself is a telling signal about where the frontier of the possible currently lies.

A separate thread of discussion is the model's safety and "transparency." OpenAI acknowledged that in specialized tests Astra has become harder to control: in its chain of reasoning it sometimes reveals less than earlier versions did. The public version, meanwhile, flatly refuses to write exploits, and the advanced cyber capabilities are locked behind a separate program for vetted security professionals. In other words, the sharpest part of its capabilities is deliberately kept away from the ordinary user.

So is it AGI or not?

Brockman's line about the "age of AGI" is a company position, not a measured fact, and it should be read exactly that way. He himself carefully hedged: the term is fuzzy, and everyone decides for themselves. By independent numbers, the model doesn't reach "human-level intelligence across the board" — it's strong in a specific class of tasks and ordinary where universal superiority was once expected.

It's more honest to describe Astra not with a slogan but with a list: a breakthrough in computer use and cybersecurity, a flat result in general intelligence, a premium price, and a noticeable step toward autonomous agents. This is an important release — but on the merits, not on the marquee.

How to try GPT-6 Astra on ChatPlus

And here's the good news. OpenAI is rolling out access gradually and with caveats: in ChatGPT itself the model, under the name GPT-6 Pro, is enabled for the Pro, Business, and Enterprise tiers, while the mass-market Plus plan doesn't get it in chat yet; in the API it arrives in waves. On top of that, availability can be patchy depending on your plan and region.

We added GPT-6 Astra to ChatPlus right after launch — so you can be among the first to try OpenAI's new flagship: open a chat with GPT-6 Astra. The model is available on the Max plan — and that's deliberate: Astra is expensive to run, so we placed it in the top tier alongside our other most capable models. It's also easy to compare it with GPT-5.6, Claude, and Gemini in one window — switch models right in the conversation and see which one handles your task best.

In short

GPT-6 Astra is OpenAI's new flagship, and the main thing about it isn't that it "answers smarter" but that it "runs your computer itself": browser, forms, code, long multi-step tasks. In computer use, cybersecurity, and science the model genuinely leads (100% on ExploitBench, nearly maxing out hard math), but in general intelligence independent tests put it on par with the previous generation — at a price about 2.5 times higher. The talk of an "age of AGI" is OpenAI's position, not a measured fact. On ChatPlus, Astra is already available on the Max plan — so you can be among the first to try it.