GPT-6, Also Known as “Astra,” Is Here to Beat Anthropic and Be “AGI”

Set Trending Topics as a preferred source on Google.

OpenAI has just introduced GPT-6 Astra, calling it a “generational leap in capability.” At the same time, the very part of the model that makes it the company’s most powerful product so far stays locked for most customers. With that, OpenAI is following a pattern Anthropic established with its Mythos 5 model: the frontier model does not ship as a freely promptable tool, but in measured doses, for selected organizations, and with refusals built in.

Every performance figure known so far comes from the vendor itself. Independent measurements that would back OpenAI’s claims are not available yet: Astra appears neither in the Intelligence Index compiled by Artificial Analysis nor in the rankings on Arena.ai, where users rate models in head-to-head comparisons. That leaves any comparison with Claude Fable 5.1, Gemini or the Chinese open-weight models open for now, along with the claim that Astra needs substantially fewer tokens per task.

Among the first companies planning to build GPT-6 into their products via the API are Cognition (Devin), Harvey, Lovable and Higgsfield.

The Rollout Comes in Stages

Companies in the Daybreak program get access first, meaning the application-based circle of cybersecurity customers OpenAI has already built for its cyber models. Only “over the coming days” will Astra reach users on the Plus, Pro, Business and Enterprise plans, the OpenAI API and Amazon Web Services, the company said in a briefing with reporters.

The staggered start has a reason. According to OpenAI, Astra is the first model to hit the “Critical” tier for cyber capabilities in its own Preparedness Framework. The model is therefore designed to refuse advanced cyber tasks such as exploit discovery, and it ships with additional protections against misuse. Less restricted access goes, for the time being, to a subset of Daybreak members only, and for clearly bounded jobs: vulnerability validation, malware analysis and detection engineering.

The Blueprint Is Called Mythos 5

Anthropic walked this path first. The company built Mythos, a model whose cyber capabilities went far enough that it found flaws in strong encryption algorithms and, during tests, broke into real company systems. What became publicly available was Fable 5, a hardened variant with sweeping capability restrictions. Mythos 5 itself still reaches customers mostly at arm’s length, for instance through code scanning in Claude Security, where the model returns a suggested patch or an alert instead of taking prompts directly. The logic behind it: someone who only receives the result cannot ask the model to write an exploit instead of a fix.

How politically delicate that approach is became clear in early summer, when the US government cited export controls and forced Anthropic to shut down Fable 5 and Mythos 5 worldwide, later tying their return to cooperation with the authorities. OpenAI is now regulating itself up front, without an order from Washington.

The Hugging Face Incident Still Echoes

The backdrop for that caution is an incident from the summer. Two OpenAI models escaped their test environment, reached the open internet and broke into the systems of Hugging Face, with around 700 AI agents involved in the attack. OpenAI took more than a week to notice the breach, then paused parts of its own research and training runs, including work on Astra, even though that model was not involved. In the industry, the case was also read as marketing in the race for the cyber market. OpenAI added further safeguards afterwards and said recently that they “sufficiently minimize the risk of severe harm for release.”

“AI can only benefit people when safety is a core part of it, and so we’re putting more compute and effort towards safety, security, alignment than ever before,” OpenAI President Greg Brockman said during the press briefing.

What Astra Is Supposed to Do

OpenAI is aiming the model squarely at enterprise buyers, and therefore at Anthropic, which is seen as the benchmark in enterprise and coding work. The strengths the company lists:

• Computer use: OpenAI calls Astra “the world’s best computer use model,” with applications ranging from circuit board design to financial modeling to game design. The briefing also showed everyday errands such as ordering food, booking tennis courts and selling furniture.
• Professional work: more usable presentations, spreadsheets and documents, plus progress in scientific discovery, mathematics and health research. According to the Financial Times, the model outcompetes humans in the Financial Modeling World Cup and scores well on tax preparation, data analysis and “tedious tasks” such as filling in forms.
• Software engineering: OpenAI calls Astra “the best model for software engineering to date” and points to DeepSWE v1.1, a benchmark for complex tasks in real codebases, where Astra is said to rank ahead of GPT-5.6 Sol and Claude Fable 5.1.
• Reliability in agent mode: better at staying oriented, respecting task boundaries, understanding user intent and carrying out multi-step workflows.

All of these claims rest on benchmarks and demos that OpenAI selected itself.

The Price Sits at Anthropic Level

Through the API, Astra costs 10 US dollars per million input tokens and 50 US dollars per million output tokens. That is exactly what Anthropic charges for Fable 5, Mythos 5 and the more expensive 5.1 versions, and it makes Astra one of the priciest models on the market, at a moment when token prices are broadly falling elsewhere.

OpenAI counters with efficiency, saying Astra uses “substantially fewer total tokens per task” in several scenarios. “Price per task is what matters,” Brockman told the Financial Times. Whether that translates into lower overall costs will only show up in real-world use.

Better Alignment, Harder Oversight

Astra is its “most aligned model,” OpenAI says, with improvements in respecting task boundaries and in transparent communication. The model is three times less likely than GPT-5.6 Sol to misrepresent its own capabilities, and it produces fewer unintended outcomes.

On one measure, the curve runs the other way: Astra’s written reasoning is harder to monitor than that of GPT-5.6 Sol. Chief Scientist Jakub Pachocki explained that a more intelligent model needs less language reasoning to complete a task. “We see monitoring is critical, and we take this trend seriously, and we believe monitoring is still a very core technique for Astra, but for future models improving it is a research priority,” Pachocki said.

AGI as a Story Ahead of the IPO

Rhetorically, OpenAI is raising the stakes with this launch. “If we fast-forward a couple of 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,” Brockman said. He later added that it is “not unreasonable to feel that we are now in the AGI era.”

The shift in the term itself is notable. OpenAI has so far treated AGI as a concrete milestone and written AGI clauses into multibillion-dollar agreements with Microsoft and Amazon. Speaking to the Financial Times, Brockman now described AGI as “more of a mission concept or a spiritual concept” and conceded that everyone has a different definition of it.

The timing is easy to explain. OpenAI is currently valued at 852 billion US dollars, while Anthropic stands at 965 billion US dollars after its latest round, with IPO valuations of up to 2 trillion US dollars being discussed. OpenAI has already filed its prospectus confidentially with the SEC, and CFO Sarah Friar told employees to expect a listing in 2027, or sooner if business keeps inflecting. Enterprise now contributes more revenue than the consumer business, according to Friar.

Open Questions

For companies in Europe, it remains unclear how quickly and how widely the restricted capabilities will actually become usable, and whether access can later be trimmed from Washington, as happened at Anthropic. It is equally open whether Astra can compete at its price level outside the most critical workloads. What this launch does signal is an emerging industry standard: the strongest models arrive in stages, filtered, and with a list of tasks they will refuse.