OpenAI's flagship model, introduced on 3 September 2026. Its most important new ability is that in ChatGPT Work and Codex it operates computer programs itself and carries multi-step tasks through on its own. It is available on the paid ChatGPT plans (Plus, Pro, Business, Enterprise) and through the developer API.

OpenAI's president says it is not unreasonable to feel that we are now in the AGI era, but the official announcement does not declare Astra to be AGI. Independent researchers point out that model performance is still uneven. For a business, the more useful question is which tasks can be handed to it reliably.

Which is better: GPT-6 Astra or Claude Fable 5.1?

In hands-on tests Astra is the more reliable executor: its features work, it installs and tests what it builds, and in business analysis it chases down missing data. Fable 5.1 delivers more visually polished work. Their developer prices are the same, so the task should decide which one to use.

Should you use GPT-6 Astra for every task?

No. Astra burns through usage quickly, so routine tasks such as a daily inbox check or news monitoring can stay on a cheaper model like GPT-5.6. Astra pays off where several steps need to be connected independently and with judgment.

GPT‑6 Astra: is AGI really here, and what does it mean for businesses?

AI news 15 September 2026 · dropmind editorial

On 3 September OpenAI released its new flagship model, GPT‑6 Astra. Before the launch, the company's president told reporters it is not unreasonable to feel that we are now in the AGI era. That is a big claim, and the point is easy to lose in the hype around it. So this article takes it apart in plain terms: what AGI is, what actually changed, and what all of this means for a company that serves customers rather than running a research lab.

Released3 September 2026

Who gets itPlus · Pro · Business · Enterprise · API

Key changeit works the computer itself

AGI?a big claim, and disputed

A spiral galaxy whose arms form the numeral six, rendered as a fine print dot screen on a light cyan-mint gradient, flanked by the words GPT and Astra

What was released, and who can use it?

GPT‑6 Astra is OpenAI's most capable model. On 3 September it went first to Pro, Enterprise and Business Premium customers in ChatGPT Work and Codex, and to developers through the API. Plus and Business subscribers followed over the next few days. Usage counts against your existing subscription allowance, and once that runs out you can buy credits for more.

One important detail: on the Plus plan, Astra currently shows up in ChatGPT Work and Codex, not in the regular chat window. On Pro, Business and Enterprise it is available in chat as well, under the name GPT‑6 Pro. If you have not used ChatGPT Work yet, this article explains what it is good for. The developer price is $10 per million input tokens and $50 per million output tokens, which is where today's flagship models sit.

What is AGI, and is it really here?

AGI (artificial general intelligence) means a system that is not just good at one narrow task but matches or exceeds human performance at almost any kind of knowledge work. OpenAI's own definition describes AGI as highly autonomous systems that outperform humans at most economically valuable work.

Greg Brockman, OpenAI's president, said before the launch that he thinks it is not unreasonable to feel that we are now in the AGI era. That sentence is worth reading carefully. It does not say Astra is AGI, and neither does the official announcement. Independent researchers are more cautious: Toby Walsh, a professor at the University of New South Wales, says AI is still very jagged, and models still stumble on simple tasks.

Hands-on tests bear this out. For one tester, Astra built a game that worked flawlessly on the first try, yet on a video editing task it did worse than its predecessor, GPT‑5.6. That is why we would not argue about the label. The useful question for a business is which tasks you can hand over with confidence, and which you cannot.

The real change: it works the computer itself

Until now, AI assistants mostly told you what to do. In ChatGPT Work and Codex, Astra operates the programs itself: it clicks, types, reads and edits files, even in software that has no integration to plug into. According to OpenAI, this makes ChatGPT nearly twice as fast at using a computer as before. Testers' reports point to four things that matter for a business too.

An audit from a whole folder

One tester handed it the complete folder of a fictional plumbing business. Astra went through the files and spotted money going to an unauthorised recipient and several customers paying late.

Using programs the way a person would

It opens a video editor and places footage at the right points on the timeline. Or it clicks through a website's download flow and reports where visitors get stuck.

It tests what it builds

When it builds an app or a game, it installs and launches it, tries it out, and only hands it over after testing.

It stops when it does not know

According to early reports, when an instruction is missing the model does not guess; it pauses the work. For a business that is worth more than a confident but wrong answer.

GPT‑6 Astra or Claude Fable 5.1? The difference in practice

Two days before Astra, on 1 September, Anthropic released a new version of Claude Fable: Fable 5.1. The two models have the same developer price, so the obvious question is which one to pick. One tester gave both the same real work tasks, and the results show clearly how differently the two models behave.

Building a website, Fable 5.1 designed the better looking page, but its video pop-ups did not work and it ignored an explicit instruction. Astra's page looked modern and every button worked. Building an app, both delivered a working program, but Astra installed and launched it on its own, while with Fable the tester had to run commands and set up the development environment manually. The business audit showed the biggest gap: Fable treated the folder it was given as the whole truth, while Astra asked for the missing data and checked the email inbox to confirm that an apparently unpaid invoice was just an admin error.

Put simply: Fable 5.1 is the talented designer whose work you need to double-check, while Astra is the dependable project lead who asks, verifies and finishes the job. This is one tester's experience rather than a lab measurement, and in practice the better model is decided task by task. In copywriting, for example, neither produced good results without samples: both needed to see earlier writing before they sounded human.

What to watch out for

Cost. Astra is powerful, but it burns through usage quickly: one tester used roughly $1,500 worth of credits in 100 hours. Another tester therefore kept routine jobs, such as a daily inbox check and news monitoring, on GPT‑5.6, because Astra's improvement there was not worth the price. A good rule: run the same task on both models and only switch where you can see the difference.

Working the screen is not always the best route. If a system has a proper integration (an API), going through it is faster, cheaper and more reliable than having the model click around on screen. Computer use is useful where no such integration exists.

Access. According to OpenAI, Astra's cybersecurity capabilities reach the highest, "Critical" level of the company's own risk framework, so it was released with stricter safeguards. For a business the practical message is simple: think through which folders, accounts and programs you give it access to, and keep a human signing off on the output.

What does this mean for your business?

With Astra, the role of AI shifts from an assistant that answers questions to a colleague that carries a multi-step task through. That moves the bottleneck too: less and less does the model's knowledge decide how much value it delivers, and more and more the company's processes, the access it grants and the quality of review.

We suggest starting with one or two multi-step processes where today a person has to connect the steps across several programs, such as putting together a monthly report. Let a cheaper model handle the routine, Astra or Fable 5.1 the steps that need judgment, and a human the final sign-off. In this market a new model arrives every few weeks, so in a good system the models are swappable and the one that best fits the task does the work.

Curious which of your processes you could already hand to an AI that works on its own, and which you should not? Let's find out together on a free 30-minute consultation.

Free 30-minute consultation

Frequently asked questions

What is GPT‑6 Astra?

OpenAI's flagship model, introduced on 3 September, which operates programs itself in ChatGPT Work and Codex and carries multi-step tasks through on its own. It is available on the paid ChatGPT plans and through the API.

Is GPT‑6 Astra AGI?

OpenAI's president says it is not unreasonable to feel that way, but the official announcement does not claim it, and independent researchers say models are still uneven. The more useful question for a business is what it can reliably hand over.

Which is better: Astra or Claude Fable 5.1?

In hands-on tests Astra is the more reliable executor and Fable 5.1 the more visually polished. Their developer prices are the same, so the task should decide.

Should you use Astra for every task?

No. It burns through usage quickly, so routine tasks can stay on a cheaper model such as GPT‑5.6. Astra pays off where several steps need to be connected independently and with judgment.

Sources

Last updated: 15 September 2026