OpenAI Launches GPT-6 Astra, Mentions AGI and Charges $10 per Million Input Tokens
The model debuted on September 3 with a phased rollout, pricing structure identical to Claude Fable 5.1, and a record in benchmarks that Brockman himself referred to as a generational leap.
OpenAI launched GPT-6 Astra on September 3, with a promise not made since the debut of GPT-4: that the model could, at some point, be viewed as the milestone for achieving artificial general intelligence. Greg Brockman, the company's president, called the release a "generational leap" and said that Astra is the result of "years of research and high-stakes investments." The rollout began with a limited set of organizations in the company's cybersecurity program, followed by a broader release to ChatGPT Plus, Pro, Business, and Enterprise customers, as well as the API and AWS over the coming days.
The public pricing for the API is set at $10 per million input tokens and $50 per million output tokens in Standard mode, with a Fast mode that doubles both speed and price. This is the same range charged by Anthropic for Claude Fable 5.1, which was launched two days earlier, and about 2.5 times the promotional price of GPT-5 Sol. The convergence of pricing between the two leading labs effectively ends the aggressive dumping cycle that dominated the first half of the year.
Saturated Benchmarks and the Evaluator's Caveat
OpenAI claims 97.6% in FrontierMath Tier 4, 99.9% in ARC-AGI-3, and 100% in ExploitBench. These numbers matter because both tests were designed to stay ahead of the capabilities of frontier models: saturating them signals something qualitatively different from merely surpassing a conventional leaderboard. However, the interpretation comes with an asterisk. Epoch AI, which maintains FrontierMath, reported that OpenAI funded the development of the test and has exclusive access to some of the problems. The 99.9% score in ARC-AGI-3 relies on a costly state harness; stateless calls via the API yield lower scores, around 98.6% according to different configurations.
Without these nuances, the release reads like a race won. With them, the picture changes: Astra leads the field in reasoning, software engineering, computer use, and cybersecurity, but the "almost perfect" must be understood within the limits imposed by the evaluators themselves. Al Jazeera characterized the launch as an event "amid growing scrutiny," referring to recent alerts from European and American regulators.
Daybreak Enters the Scene as Well
On the same day, OpenAI launched the Daybreak for Frontline Defenders program, with $1 billion in subsidized credits for access to Daybreak cybersecurity models, training, and partnerships aimed at essential service operators. The simultaneous announcement is no coincidence: Astra is described as state-of-the-art in cybersecurity, and the company knows that putting a model capable of discovering zero-days into the hands of any paying customer invites immediate regulatory scrutiny. Tying the release to a defensive initiative is a structured response.
Considerations for Corporate Clients in the U.S. and U.K.
For the North American CIO, the practical decision is whether to replace Claude Fable 5.1 with Astra in code pipelines, given that the pricing is the same. The answer depends on the relative weight of mathematical reasoning (where Astra excelled) versus agent-to-agent latency and tool usage (where Anthropic has a longer track record). In the U.K., the Financial Conduct Authority has already indicated that it will require evidence of independent evaluation before approving frontier deployments in regulated banks, and OpenAI's partial exclusivity on FrontierMath is likely to become ammunition for compliance officers to call for outsourcing benchmarks.
There is a more immediate secondary effect: Google's Gemini 3.1 Pro, launched in February with a context of 1 million tokens, is no longer the model with the best agent appeal in various evaluations. The window in which Sundar Pichai could argue leadership in reasoning benchmarks has closed. Google's next move, expected before the end of the quarter, will determine whether the triple parity among OpenAI, Anthropic, and Google consolidates or whether one of the three stands out again in price, and not just in capability.