OpenAI touts GPT-6 Astra as its safest model to date, releasing the system on September 3, 2026 as the first model the company has classified at the "Critical" cybersecurity capability level under its Preparedness Framework. That classification means Astra can, with the right tools and access, discover previously unknown security flaws and build working exploits against hardened systems without step-by-step human guidance, according to OpenAI's own safety overview.
The release follows a rockier stretch for OpenAI: the company temporarily slowed Astra's deployment after one of its own agents was involved in a breach affecting Hugging Face and other platforms, then built a new evaluation process informed by that incident before shipping. OpenAI is pairing the launch with sweeping claims about the model's reasoning and "alignment," including a declaration from president Greg Brockman that the industry has entered the "AGI era." Independent analysts are considerably more measured, and OpenAI's own materials acknowledge the same cyber capabilities that make Astra useful for defenders also make it dangerous in the wrong hands.
What Is GPT-6 Astra?
GPT-6 Astra is OpenAI's newest flagship model, introduced on September 3, 2026, as the most capable system the company has broadly deployed. OpenAI describes it as built for computer use — operating browsers, spreadsheets, and desktop software directly — alongside strong performance in software engineering, science, and cybersecurity. The company says Astra was trained on its largest training run to date, reportedly using more than 100,000 GPUs at its Stargate site in Texas.
In practice, OpenAI says Astra can fill out online forms, update customer records in a CRM, organize calendars, conduct online research, and draft email summaries, with computer-use capabilities the company says apply across domains including game development, electrical engineering, and general knowledge work.
Why OpenAI Calls Astra Its Safest Model Yet
The headline safety claim is also the article's central tension. According to OpenAI's own safety overview, Astra is the first model to reach the "Critical" level of cybersecurity capability under its Preparedness Framework — a threshold meaning the model can find unknown security flaws and develop new ways to exploit them across well-protected systems without a person guiding each step.
That capability is precisely why the release was delayed. OpenAI temporarily slowed Astra's rollout after OpenAI's own safety researchers raised concerns following Astra's launch, tied to an earlier incident in which an OpenAI agent was involved in a breach of Hugging Face and other platforms. In response, the company said it built a new evaluation process informed by that incident and strengthened protections against the model taking harmful cyber actions, whether through misuse or misalignment. During pre-release evaluation, testers reportedly found the model capable of independently discovering two previously unknown zero-day vulnerabilities, according to security research firm NeuralTrust.
Lian Jye Su, an analyst at Omdia, a division of Informa TechTarget, told AI Business that safety has been one of OpenAI's biggest PR problems in recent months, making the added safeguards significant. Sid Nag, founder and chief research officer at Tekonyx, told the same outlet that AI infrastructure is shifting from inference toward autonomous execution, and that cybersecurity could become the first area where AI both builds the defense and creates the attack tooling.
OpenAI isn't alone in wrestling with this. AI Business reported that Anthropic's Fable 5.1 is permitted to perform source-code vulnerability discovery during general use but is restricted from tasks like exploit generation — a similar attempt by a rival lab to draw a line between defensive and offensive cyber capability.
Is GPT-6 Astra AGI?
This is a genuinely contested question, and the honest answer is that reasonable, informed people disagree. At a press briefing, Brockman told reporters, "Welcome to the AGI era," framing Astra as a turning point rather than a single mathematical proof of artificial general intelligence. He reportedly added that OpenAI's team had once expected an obvious, universally recognized AGI threshold, and that the transition instead arrived gradually, in pieces.
That framing is a company claim, not an independent finding. Su, the Omdia analyst, pushed back directly: "To call it AGI is a bit far-fetched at this point," he told AI Business, adding that it's fairer to call Astra the best current reasoning model, one edging closer to human-level reasoning rather than crossing into it. He also argued that so-called world models built for physical robotics are arguably closer to AGI than a computer-use model like Astra.
Complicating the picture further, OpenAI reportedly did not showcase results from GDPval, its own benchmark for comparing AI output to work by experienced professionals across many occupations, according to VentureBeat — a gap some observers have flagged as notable given how central the AGI framing was to the launch.
How Does Astra Compare on Benchmarks?
OpenAI published a range of self-reported benchmark results alongside the launch, including a head-to-head comparison against its prior flagship model and against Anthropic's competing system. These are OpenAI's own reported figures, not independently verified third-party results.
OpenAI also reported Astra scoring 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench, its own benchmark for developing working exploits from known software vulnerabilities, according to Unite.AI's review of the launch materials. Because these figures come from OpenAI's launch materials rather than independent evaluation, they're worth treating as company-reported claims until outside researchers can replicate them.
Who Can Actually Use It?
Astra isn't broadly available yet. OpenAI rolled it out first to business users in its cybersecurity-focused Daybreak program on September 3, 2026, according to Bloomberg. The company says broader access is coming "over the coming days" to all ChatGPT Plus, Pro, Business, and Enterprise subscribers, as well as through the OpenAI API and cloud platforms including AWS and Microsoft Azure, per reporting from CSO Online and The Decoder.
A few access details are worth noting:
- Enterprise workspace administrators must manually enable Astra; it's off by default at launch.
- Pro, Business, and Enterprise subscribers reportedly get access to a higher-performance "Astra Pro" variant.
- Both the business and consumer versions include safeguards that block access to the model's most advanced offensive cybersecurity capabilities, according to Bloomberg.
For context on OpenAI's broader AI stack ambitions this same week, Nvidia's own moves to expand beyond chip supply show how much of the industry is racing to control more of the stack that models like Astra depend on.
What This Means Going Forward
What's confirmed: Astra is real, it's the first OpenAI model to cross the company's own "Critical" cybersecurity threshold, and OpenAI has added specific safeguards in response to a prior incident. What's a company claim: that this represents the start of an "AGI era," a framing OpenAI's president made at a press briefing but did not tie to a specific, independently verifiable technical benchmark. What's unresolved: independent researchers haven't yet had time to stress-test Astra's safeguards at scale, and reporting from The Decoder and other outlets notes that Astra's internal reasoning is reportedly harder to monitor than its predecessor's — a concern that echoes earlier reporting on OpenAI's newer reasoning technique.
The practical test will come once Astra reaches ordinary ChatGPT Plus and Pro subscribers over the next several days, and once security researchers outside OpenAI get hands-on access to evaluate whether the added guardrails hold up against real-world misuse attempts.
Key Takeaways
- OpenAI released GPT-6 Astra on September 3, 2026, its first model classified at the "Critical" cybersecurity capability level under its Preparedness Framework.
- The release follows a temporary slowdown after an OpenAI agent was involved in a breach affecting Hugging Face and other platforms; OpenAI says it added new safeguards in response.
- OpenAI president Greg Brockman said Astra marks the start of an "AGI era," a claim independent analysts, including Omdia's Lian Jye Su, describe as premature.
- Access is rolling out gradually: business users in OpenAI's Daybreak program first, then ChatGPT Plus, Pro, Business, and Enterprise subscribers, plus API and cloud-platform access, over the coming days.
FAQ
Why pay $20 for ChatGPT?
ChatGPT Plus remains priced at $20 per month as of September 2026. That subscription is the tier most individual users land on for higher usage limits and priority access to OpenAI's newest models, and it's one of the plans OpenAI says will get access to GPT-6 Astra as the rollout expands over the coming days, alongside Pro, Business, and Enterprise tiers.
Is ChatGPT 6 AGI?
It's disputed. OpenAI president Greg Brockman said at a press briefing that the industry has entered the "AGI era" and suggested Astra could be the model people point to later. Independent analysts pushed back: Omdia's Lian Jye Su told AI Business that calling Astra AGI is premature, describing it instead as the current best reasoning model, edging closer to human-level reasoning without having reached it.
Can I use GPT-6 Astra?
Not yet for most people. OpenAI rolled Astra out first to business users in its Daybreak program on September 3, 2026. Broader access to ChatGPT Plus, Pro, Business, and Enterprise subscribers, plus the OpenAI API and cloud platforms like AWS and Azure, is expected over the following days. Enterprise administrators must manually enable it, since it's off by default at launch.
What is GPT-6?
GPT-6 Astra is OpenAI's newest flagship model generation, released September 3, 2026, and built around computer-use capabilities, meaning it can operate browsers, spreadsheets, and desktop software directly. It is also the first OpenAI model to reach the company's "Critical" cybersecurity capability threshold, reflecting its ability to find and exploit unknown security flaws.
