New Delhi: GPT‑6 Astra, OpenAI’s new flagship artificial intelligence model, began its rollout Friday in what is being billed as the advent of a new AGI (Artificial General Intelligence) era.
OpenAI describes Astra, which succeeds GPT‑5.6 Sol, as its model for computer use, browsing, software engineering, cybersecurity, science and professional work.
It is trained to operate software on screen the way a person does—filling out online forms, updating customer records, organising calendars, conducting research, drafting summaries, and building and hosting websites. OpenAI lists financial modelling, tax preparation and game development among example tasks.
At its launch, company president Greg Brockman told reporters that the release could later be seen as the point at which artificial general intelligence arrived. “I think it might be about this model,” he said, ending the briefing with the words: “Welcome to the AGI era.”
OpenAI’s charter defines AGI as systems that outperform humans at most economically valuable work, a threshold written into its agreements with Microsoft and Amazon.
Brockman told the Financial Times that he now viewed AGI as a “mission concept,” and said definitions differ between people. Altman had earlier described the term as poorly defined.
Chief executive Sam Altman told CNBC that the model represented a new capability level and had changed his own workflows.
OpenAI says the model came from its largest training run, using more than 100,000 GPUs at its Stargate site in Texas, and that it was the first time the company used other models to supervise training.
Also Read: OpenAI knew its AI models were behaving unusually before they went rogue & hacked Hugging Face
What does OpenAI claim it can do?
On benchmarks published by the company, Astra scored 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 57.9% on Terminal-Bench 4.0, a software-engineering test. On the OSWorld 2.0 computer-use test, OpenAI reports Astra scored 72.6% at about 40 minutes per task, against 65.7% at about 75 minutes for the previous model.
Astra, according to OpenAI, contributed to two mathematical results on gaps between prime numbers, including improving a bound (the maximum and minimum values a number could have been before it was rounded) it says stood for more than 80 years.
Independent measurement released within hours told a different story. Artificial Analysis, a firm that benchmarks models for the industry, reported that Astra ranked level with OpenAI’s previous generation and behind the leading models from Anthropic and Meta’s Muse Spark 1.3 on its two headline indices.
OpenAI’s own figures show a similar pattern on some measures. On the Artificial Analysis Intelligence Index, Astra scored 61.2, against 65.7 for Anthropic’s Claude Fable 5.1 and 63.1 for Claude Opus 5. On the same firm’s coding index, Astra scored 67.0, close to Anthropic’s models. Google’s Gemini 3.8 Flash trailed on most listed tests.
The cybersecurity warning
OpenAI says Astra is its first model to reach the “Critical” threshold under its Preparedness Framework — a level reached when a model can find and build working zero-day exploits against protected systems without a person directing each step.
During internal testing, it says, Astra found and used two previously unknown vulnerabilities, which it is disclosing to the affected maintainers. On the ExploitBench test, run without production safeguards, OpenAI reports Astra scored 100%, against 78.5% for the prior model.
The new version refuses tasks such as writing proof-of-concept exploits. OpenAI says its tests found Astra’s written reasoning harder to monitor than the previous model’s when it was instructed to evade oversight, and that it is deploying classifiers in production to check the reasoning and actions of such models and halt unauthorised activity. Those checks, it notes, can interrupt legitimate work.
How access is being rolled out
Companies in OpenAI’s application-based Daybreak programme get access first, with wider release to ChatGPT Plus, Pro, Business and Enterprise tiers over the following days, along with the API and Amazon Web Services.
In the API, Astra is priced at $10 per million input tokens and $50 per million output tokens, with a faster mode at twice the speed and twice the price.
Altman said the release was delayed. OpenAI slowed the model in August because tests could not rule out a ‘Critical’ risk level. He also said the model went through a review by the Trump administration before its release.
Astra is positioned against Anthropic, seen as a reference point in enterprise and coding work, and against Google’s Gemini line and Meta’s Muse Spark. OpenAI is valued at about $852 billion. Anthropic stands at about $965 billion after its latest funding round, with figures near $2 trillion cited in discussions of a possible public listing.
(Edited by Tony Rai)