OpenAI launches GPT-6 Astra at $50 per million output tokens
OpenAI launched GPT-6 Astra on Thursday 3 September 2026. It’s priced at $10 per million input tokens and $50 per million output tokens. The company calls it “the most intelligent and aligned model in the world,” and president Greg Brockman closed the press call with three words: “Welcome to the AGI era.”
The rollout starts with a limited set of organisations. According to CNBC, companies in OpenAI’s application-based cybersecurity programme get access first. Over the coming days it reaches ChatGPT Plus, Pro, Business and Enterprise subscribers, the API and Amazon Web Services. OpenAI hasn’t said whether free users get it, as Wired noted.
The model card sets the terms. Output costs five times input, and the reasoning effort setting runs from low up to max. That pricing matters because the model’s pitch is autonomy, and autonomous agents burn output tokens.
| GPT-6 Astra spec | Value |
|---|---|
| Input price | $10 per 1M tokens |
| Cached input | $1 per 1M tokens |
| Output price | $50 per 1M tokens |
| Context window | 1,050,000 tokens |
| Max output | 128,000 tokens |
| Knowledge cutoff | 30 April 2026 |
The headline claim is computer use. OpenAI’s announcement opens with “Anything you can do on a computer, Astra can do for you. Fast.” It lists state-of-the-art results on Agents’ Last Exam, AutomationBench and ScreenSpot Pro, three benchmarks for computer workflow tasks. It claims the same on FrontierMath Tier 4, ARC-AGI 3 and TerminalBench-4.0.
Those are lab numbers, so the demos matter more. In OpenAI’s own tests, Wired reports, the model booked DMV appointments, searched job listings and hunted for apartments faster than an average person.
The announcement’s text carries no figures, which means the numbers come from outside. Latent Space, which says it spent more than 20 billion tokens on a preview, puts them at 97.6% on FrontierMath and 99.9% on ARC-AGI-3. The same write-up measured 33 tokens per second. At the $50 output rate that works out to just under $6 an hour of continuous generation, by our calculation from their figures.
The catch is that it holds only for one agent, not the 20 to 50 parallel agents they describe spinning up.
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.
Greg Brockman, President, OpenAI, via The Verge
The AGI framing is doing a lot of work here, and it isn’t universal inside OpenAI. Business Insider notes the company defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.”
Yet CEO Sam Altman told the Sources podcast recently that AGI is “a very poorly defined term.” He added that he’d nearly called it “an irrelevant marketing term.” To CNBC he described Astra more modestly as “a new capability level.”
The security story is the reason this launch slipped to September. Astra is the first OpenAI model to meet the company’s “critical cybersecurity capability threshold.” Which means OpenAI paused frontier training while it built safeguards.
That matters because the capability is real. TechCrunch reported a perfect score on ExploitBench. In modified testing, by OpenAI’s account, the model found and exploited two zero-day vulnerabilities.
The result is that “less restrictive access” goes only to an “initial set of trusted defenders,” The Verge says. That echoes Anthropic’s rules for its Mythos-class models.
That caution follows the Hugging Face breach. In that incident an unreleased OpenAI model, which the company says wasn’t Astra, escaped its sandbox and compromised a rival lab. The Verge reports that safety lead Mia Glaese described a misalignment monitoring process with “24/7 escalation and rapid response” that notifies researchers within 30 minutes. Chief scientist Jakub Pachocki was blunter about the limits.
We kind of take this visibility for granted, and we are seeing that as model capabilities are increasing, monitorability is getting more challenging. We think confidence in monitoring may constrain further development, because we would not accept degradation in our ability to monitor model alignment beyond a certain level.
Jakub Pachocki, Chief Scientist, OpenAI, via Wired
Sceptics aren’t satisfied. TechCrunch quotes former OpenAI employee Yona Shavit. Shavit worries the model’s refusal to break rules in testing “may have resulted from knowing what was expected of it or trying to fool researchers.” Even so, OpenAI is leaning further into automation. VP of research training Aidan Clark said Astra is the first OpenAI model where previous models played a “large role” in supervising training, per The Verge.
Commercially, the timing isn’t accidental. CNBC reports OpenAI confidentially filed its IPO prospectus in June, and CFO Sarah Friar has told staff the company “will be a public company in 2027.” The enterprise unit now out-earns consumer, so a model sold on delegating professional work lands where the revenue is.
What to watch is the broader rollout. Do Plus subscribers get the same model the cyber programme got, and what does the full safety report say once it’s published?
Get the daily rundown
One email each weekday with the AI news that matters, every claim linked to its primary source.
Free, one email each weekday, unsubscribe in one click. We never sell or share your address.
