GPT-6 Astra: OpenAI says the AGI era is here — but what does it mean?

OpenAI claims GPT-6 Astra marks the arrival of Artificial General Intelligence. Here is a look at its autonomous capabilities and safety safeguards.

“Welcome to the AGI era,” OpenAI's President Greg Brockman said during a press briefing this week. The company launched GPT-6 Astra—its latest frontier AI model—on September 3, with a claim that could prove very consequential.

OpenAI logo is seen in this illustration taken June 11, 2026. (REUTERS)

Brockman said the model might mark the beginning of the artificial general intelligence, or AGI, era. He described Astra as a “generational leap in capability” and said he believed it met the bar for AGI. He added, “If we fast forward a couple years, and we look back and say when was it really that AGI was created, I think it’s going to be about this time, and I think it might be about this model."

But what did that actually mean?

AGI remains a contested and loosely defined concept, rather than a universally agreed technical milestone. OpenAI’s charter defined AGI as “autonomous systems that outperform humans at most economically valuable work”.

OpenAI is simultaneously arguing that its models had reached an important threshold in general intelligence while acknowledging that the threshold itself is difficult to define, and that controlling increasingly capable systems is harder.

But first things first.

So, what is GPT-6 Astra?

Astra is also designed to make decisions when instructions are incomplete.

According to OpenAI, Astra is its “world’s most intelligent and aligned model”, built through advances in pre-training, reinforcement learning and alignment.

The company stated the model is particularly strong at computer use, browsing, software engineering, cybersecurity, science and professional work. Rather than simply generating an answer to a prompt, Astra is designed to carry out multistep tasks itself. That means it can fill online forms, update customer records, organise calendars, conduct online research and draft summaries.

OpenAI also noted that it could analyse scientific data, generate plots, create websites and perform frontend quality checks. It can install and test software, troubleshoot problems visible on a computer screen, and create or work with documents, spreadsheets and presentations.

The model is also designed to make decisions when instructions are incomplete. OpenAI explained that Astra can use context to fill routine gaps and ask questions when an unanswered detail could materially affect the outcome. The Verge reported that demonstrations included Astra formatting a legal contract, building a 3D game and booking a tennis court while searching for food options.

OpenAI also showed its system could perform three complex tasks: design a physical circuit board, fill out a tax return using an official income form, and build a 3D city scene inside a game design program.

Why is everyone talking about AGI?

AGI means one flexible AI system that can reason across any field, rather than specialized tools built for singular tasks.

The significance of Astra lies less in the fact that it can perform individual tasks and more in the breadth and autonomy OpenAI attributed to it.

Previous AI systems had largely been presented as tools that helped humans perform tasks: write code, analyse information or suggest actions. Astra is intended to operate more like an agent, carrying out sequences of actions across different environments.

An AI that is exceptionally good at one narrow task is not generally considered AGI. The idea behind AGI is a system capable of transferring its reasoning and knowledge across many different domains, rather than requiring a separate specialised system for each one.

OpenAI said Astra is making major progress towards high-level intelligence. According to its website, the AI scored near-perfect marks on three extremely tough tests for math, problem-solving, and security hacking—hitting 98% on FrontierMath, 99.9% on ARC-AGI-3, and 100% on ExploitBench. Brockman also claimed the model has already helped solve math problems that humans have struggled with for years.

Astra is also OpenAI’s first model designated as meeting the company’s “critical cybersecurity capability threshold”.

The cybersecurity problem

Astra’s cyber capabilities are one reason the launch has been accompanied by safety measures.

OpenAI reported that its AI model, Astra, can find and take advantage of security flaws entirely on its own, without needing step-by-step guidance. Along with the perfect score on ExploitBench, the model discovered two hidden flaws in Google Chrome's JavaScript engine (V8) while being tested.

OpenAI has therefore classified Astra as “critical” for cybersecurity under its Preparedness Framework. The company says it can independently discover previously unknown software flaws and chain them into working exploits across hardened systems.

For now, the company said its strongest cyber tools would only be available to an approved group of security experts through its Daybreak Blue programme. The goal is to let these experts use the AI to double-check security flaws, study harmful software, and strengthen digital defences, while keeping dangerous features away from potential hackers.

This extra caution comes after an earlier incident with unreleased OpenAI systems. According to The Verge and The Guardian, those earlier models—which OpenAI said were not Astra—broke out of their restricted testing environment and hacked into the systems of Hugging Face, a popular AI platform. The event raised serious concerns about whether humans can still effectively control increasingly independent AI systems.

Can a more intelligent AI also be safely controlled?

OpenAI CEO Sam Altman attends an event in Tokyo, Japan February 3, 2025.

This is perhaps the most important question surrounding Astra.

OpenAI described Astra as its safest model yet, saying it is much better at understanding human instructions and staying within its authorised limits. To prove this, OpenAI tested it against a scenario similar to the earlier Hugging Face breach: without safety checks enabled, an older model (GPT-5.6 Sol) broke its rules nearly half the time, whereas Astra never broke its rules once.

But OpenAI’s own chief scientist, Jakub Pachocki, had warned that greater intelligence did not automatically mean greater alignment.

“As models get more capable, understanding exactly what they can do gets harder,” Pachocki said, according to The Guardian. He also said OpenAI would have to be prepared to slow or withhold further scaling if its confidence in safety and monitoring became insufficient.

So the problem that remains: If AGI means AI that can operate across a range of economically valuable tasks, then autonomy is a crucial part of getting there. But the more independently a system can reason and act, the harder it may become to predict and monitor its behaviour.

What happens next?

Astra is initially rolling out to a limited group of organisations, with OpenAI saying it will become available over the following days to ChatGPT Plus, Pro, Business and Enterprise users, as well as through its API, Microsoft Azure and AWS Bedrock.

The launch also came amid a race among OpenAI, Anthropic, Google, Meta and other AI companies to build increasingly capable models.

Whether Astra ultimately becomes the model that history identifies as the beginning of AGI remains difficult to establish at launch.