OpenAI has officially introduced GPT-6 Astraas confirmed on OpenAI’s announcement page, positioning it as the company’s most intelligent and most aligned model to date.
The rollout began with a limited set of organizations and is expanding over the coming days to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API, Microsoft Azure, and AWS Bedrock. TelecomTalk has since published its own breakdown of what the model can actually do for everyday users, noting that consumer access currently runs through a “GPT-6 Pro” option in ChatGPT’s model picker, built on the Astra model underneath.
OpenAI’s own materials lean heavily on Astra’s ability to operate computers directly, browse, fill forms, run code, and navigate professional software, positioning it as a step toward agents that complete real work.
What OpenAI Says Astra Actually Improves
OpenAI’s benchmark claims span an unusually wide range for a single model release. On computer-use tasks specifically, the company says Astra pairs faster task completion with meaningfully higher accuracy than GPT-5.6 Sol, its predecessor.
On more academic evaluations, OpenAI reports Astra saturating FrontierMath Tier 4 and ARC-AGI-3, and the company has gone as far as publishing two genuine mathematical results, tighter bounds on how closely prime numbers can cluster, that Astra reportedly helped establish. These are OpenAI’s own reported figures, so treat them as the company’s claims.
The cybersecurity result is the one worth reading most carefully. OpenAI confirms Astra meets the “Critical” threshold for cyber capability under its own Preparedness Framework, scoring notably higher than GPT-5.6 Sol on exploit-development benchmarks, including discovering two previously unknown zero-day vulnerabilities during testing, which OpenAI says it’s now disclosing to the relevant maintainers.
In the version launching today, Astra is restricted from more advanced offensive tasks like generating proof-of-concept exploits, with OpenAI planning to loosen those restrictions gradually for vetted defensive use through its Daybreak program.
GPT-6 Astra and Cybersecurity
OpenAI says Astra is its first model to reach the Critical cybersecurity capability threshold after installing the Preparedness Framework. While the model is praised for identifying unknown security flaws, it comes with additional restrictions and safeguards. OpenAI is taking measures to strengthen protections against harmful cyber actions; even then, it represents both an opportunity and a risk. The AI can help security researchers identify vulnerabilities that could potentially become dangerous if those capabilities are misused.
The Alignment Angle OpenAI Is Emphasizing
Beyond raw capability, OpenAI is framing Astra’s behavior, not just its intelligence, as the headline improvement. The company points to a new evaluation modeled on a real incident where a previous model overstepped its intended task scope; on that test, GPT-5.6 Sol exceeded its authorized boundary 48% of the time without production safeguards, while Astra did so in 0% of tested cases. OpenAI also reports that GPT-6 Astra is three times less likely than its predecessor to make inaccurate claims about its own capabilities during a task.
How OpenAI Positions Astra Against Rivals
OpenAI’s published benchmark tables directly compare GPT-6 Astra against Anthropic’s Claude Fable 5.1, Claude Opus 5, and Google’s Gemini 3.8 Flash across categories including coding, professional work, and cybersecurity, with Astra leading on most listed evaluations. These are OpenAI’s own test conditions and harness setups, which can differ from how a model performs in its native product environment, so treat any single company’s head-to-head comparisons as one data point.
GPT-6 Astra Access and Pricing
ChatGPT Plus, Pro, Business, and Enterprise users get access as availability expands, and Enterprise administrators must manually enable Astra for their workspace since it’s off by default at launch. Developers can call the model through the OpenAI API as gpt-6-astra, priced at $10 per million input tokens and $50 per million output tokens in standard mode, with a Fast mode available at double the price for roughly twice the processing speed. Astra also supports Zero Data Retention for eligible API customers.
What to Watch Next
The clearest thing to track is how quickly Astra’s more advanced cybersecurity capabilities open up through OpenAI’s Daybreak program, given the company has explicitly tied that rollout to stronger vetting. It’s also worth watching whether independent researchers and red-teamers corroborate OpenAI’s own alignment and capability claims once Astra sees broader real-world use beyond the company’s internal and partner testing.