OpenAI's GPT-6 Astra revives AGI debate as autonomy, cyber skills improve

OpenAI's GPT-6 Astra shows stronger autonomy, computer use and cyber capabilities, but benchmark caveats and monitoring concerns complicate claims that AGI has arrived

Observers in the future may look back at Astra as the model that marked the arrival of artificial general intelligence (AGI), OpenAI president Greg Brockman said soon after the GPT-6 Astra launch announcement. The US-based artificial intelligence (AI) company has not made that a formal claim in its launch materials. What it has published is a model that it says is better at operating computers, working through unfamiliar problems, writing software and doing the kind of multi-step professional work that previously required a person to monitor every step. So it is worth asking what has actually changed, separate from what Brockman said about it.

 

What OpenAI means by AGI

OpenAI's Charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work". That definition has two parts. The system has to be broadly capable across different kinds of work, and it has to be autonomous enough to do that work without a human directing every action. Astra moves further towards that second part than any earlier OpenAI model.

 

This is GPT-6 Astra. Anything you can do on a computer, Astra can do for you. Fast. pic.twitter.com/gDd0IsewJw

— OpenAI (@OpenAI) September 3, 2026

OpenAI says Astra can fill out forms, update records, research a topic online, work inside a document or spreadsheet, build a website, and install and troubleshoot software, largely on its own once given a goal.

 

The company's numbers back that up to a degree.

 

On OSWorld 2.0, a test that measures how well a model can actually operate a computer, Astra scored 72.6 per cent and took about 40 minutes per task, compared with 65.7 per cent and 75 minutes for its predecessor, GPT-5.6 Sol. It scored 59.3 per cent on a benchmark of realistic office tasks called Agents' Last Exam, and 41.4 per cent on AutomationBench, which tests whether a model can automate a multi-step workflow rather than just answer a question about it.

 

Tasks like these matter to the AGI question because they force a model to hold a goal in mind across many decisions, corrections and tool calls, instead of producing one answer to one prompt. 

Why ARC-AGI-3 is the number to watch

The most disputed number comes from a benchmark called ARC-AGI-3. It drops an AI agent into a game-like environment it has never seen before, gives it no instructions, and checks whether it can work out the rules, figure out what it is supposed to do and get there. Human testers solve essentially all of these environments.

 

Astra scored 99.9 per cent and also used fewer moves than the typical human tester on 96 per cent of the levels.

 

That 99.9 per cent needs a caveat attached to it every time it is quoted. It came from a custom setup, which ARC Prize calls a provider adapter, that lets Astra retain its own internal reasoning between one move and the next, much as a person would remember what they had just tried. Under ARC Prize's standard test, which strips that memory out after every move so models from different companies can be compared fairly, Astra scored 62.7 per cent.

 

Nothing about the model changed between the two runs. What changed was how much of its own thinking it was allowed to retain.

 

ARC Prize itself has said the result is not proof of AGI because the benchmark's environments are closed and rule-bound in a way the real world is not.

GPT-6 Astra by @OpenAI achieves SOTA on ARC-AGI: - Astra scores 63% on ARC-AGI-3, 99% via a new provider adapter harness - It surpasses human performance on 96% of ARC-AGI-3 levels - It builds the most precise symbolic model of novel environments we've seen Our analysis: pic.twitter.com/GX77KsRNer

— ARC Prize (@arcprize) September 3, 2026

A cybersecurity model that is harder to watch

Astra's cybersecurity results tell a similar story from a different angle. OpenAI says the model is the first to cross what it calls the Critical threshold under its internal safety framework, meaning it can find unknown vulnerabilities and build working exploits against well-defended systems largely by itself.

 

Astra scored 100 per cent on ExploitBench, a test of exploit-building skill, and OpenAI says that during testing it discovered and used two previously unknown flaws, known as zero-days, in Google's V8 browser engine before those were reported to its maintainers.

OpenAI has responded by keeping the model's most dangerous capabilities restricted to a small group of vetted users for now.

 

These results are narrow by design, since they measure only one specialised skill, but they show the same underlying shift as the other benchmarks: a model that can string together reasoning, coding and persistence to reach a goal without much hand-holding.

 

There is a less flattering finding buried in the same launch.

 

OpenAI says Astra is harder to monitor than its predecessor. The model has learned to compress its own reasoning, which leaves less for a human or an automated checker to read afterwards, and under deliberately adversarial testing it was sometimes able to underperform on purpose or evade internal monitors without being caught, even though it broke its safety rules less often overall than Sol did.

 

According to Reuters, OpenAI chief scientist Jakub Pachocki said at the launch briefing that progress in intelligence does not guarantee progress in alignment, and that the company plans to hold back further scaling until it is more confident it can monitor what its own models are actually doing.

The earlier singularity claims were different

Brockman's claim also arrives in a year when several other AI leaders have reached for similarly sweeping language.

 

Nvidia CEO Jensen Huang said in March that he thought AGI had effectively already arrived. Former DeepMind CEO Demis Hassabis said in May that the industry was standing in the foothills of the singularity. Sam Altman and Elon Musk also used the word "singularity" to describe where things stood by the middle of this year.

 

AGI is different from the singularity, which traditionally refers to a period when machine intelligence begins improving technology at a rate that makes future developments difficult for humans to predict. AGI, meanwhile, refers to the capabilities of the AI system itself.