ChatGPT Astra: what OpenAI's new model means for your work

On 3 September 2026, OpenAI released GPT-6 Astra, the model you'll meet inside ChatGPT as "Astra". The most striking part isn't the chat. Astra is built to keep working on a single task for longer: operating software, writing code, doing research and producing documents. OpenAI calls it its most capable model to be broadly rolled out, and hinted in the press briefing that you could call this "agi". That last claim is exactly what the debate is about.

Below is what's factually new, what the press wrote about it, and what we make of it now that we've worked with it.

What Astra is

Astra is a model tuned for work made up of steps rather than a single answer. The developer documentation is explicit about it: complex reasoning, code, computer use, research and drafting documents. You can set per task how deeply the model may think, from low to maximum, and the window in which it holds information is large enough for a full case file — over a million tokens, or hundreds of pages. The model's knowledge runs to the end of April 2026.

"Computer use" doesn't mean free access to your laptop. The model works inside a sandboxed environment supplied by the builder, with approval moments built in. That's an important distinction that gets lost in many headlines.

Where you'll notice it in ChatGPT

The rollout was messy. At launch, Astra first went to a handful of selected companies and inside ChatGPT's work and coding environment. Paying users followed "in the coming days", which drew plenty of complaints in the Netherlands. By now Plus and Pro subscribers have access, and it's rolling out further to Business and Enterprise.

For API pricing, OpenAI names 10 dollars per million input tokens and 50 dollars per million output tokens. How that compares to the previous flagship model is explained differently across outlets, so we won't make a hard claim on that.

The agi debate in two paragraphs

OpenAI president Greg Brockman said the model performs on a par with humans across all sorts of tasks and that you could therefore call it general intelligence. The Dutch press was sober about it. Tweakers called agi a vague term without a clear meaning, and NRC's headline welcomed us to the age of artificial general intelligence, "but nobody knows what that means".

Our own reading: the real gain sits in agentic work — operating software, terminal tasks, sustaining longer runs — not a leap in general insight. On some published benchmarks Astra still trails Claude. And it landed in a week when Anthropic, Google and Meta also released a new flagship model, which puts the superlatives into perspective.

The point that really stands out: safety

Astra is the first model to reach the "critical" level on cybersecurity capability within OpenAI's own risk framework. That means the model is capable of things that can be misused, which is why OpenAI has taken extra measures. Reuters accordingly framed the launch against growing concerns about the safety of AI agents.

For you as an organisation, that's no reason to stay away, but it is a reason to agree what an agent is allowed to do: which systems, which data, who approves, and what gets logged.

What we do with it, and what we advise

• Use it where the task is made up of steps. Reviewing a migration, cleaning a dataset, summarising a case file, searching through a codebase. For short questions, a lighter and faster model is perfectly fine.
• Set the thinking level deliberately. Everything on maximum is expensive and slow. High where it matters, low where it's routine.
• Let it propose, not execute. Especially with payments, publishing and production environments, a human keeps the final say.
• Feed it your own sources. The large context window is only useful once you put accurate documents into it. Otherwise the model fills the gaps itself.
• Measure one thing. Turnaround time, error rate, or how much you still have to rewrite yourself. Without a number it stays a feeling.
• Expect an uneven rollout. What works in one environment today may sit elsewhere tomorrow. Don't build your process around a single button in a single interface.

What this means for agents in your organisation

The practical upshot is that digital employees become a lot more useful for work that takes longer than a single question. Think of an agent that handles a request from start to finish, or one that lines up your product data and stock levels every day and only flags what doesn't add up.

Want to know where that could apply in your organisation: we've mapped out the applications under digital employees and explain the basics under AI agents. For broader context, artificial intelligence and AI and data are useful, and on the discoverability side — since people increasingly search inside a chat rather than in Google — see is SEO dead or is it changing and ChatGPT ads.

In short

Astra is a serious step forward in step-based work, not a breakthrough in reasoning. You can ignore the agi flag, but not the safety classification. And you'll only capture the gain once you pick a process, set boundaries, and measure whether it's actually getting better.

Want to think through where this could fit at your organisation? Ask your question or plan a no-obligation strategy call. We'll also just tell you if the answer is "not quite yet".