OpenAI Launches GPT-6 Astra, Says Its Most Capable Model Yet Might Be AGI

OpenAI released GPT-6 Astra on September 3, calling it the world's most intelligent and aligned model to date, with new highs on benchmarks for computer use, coding, cybersecurity and science. The model is rolling out first to a limited set of organizations, with access for ChatGPT Plus, Pro, Business and Enterprise users, and for developers through the OpenAI API, Microsoft Azure and AWS Bedrock, expected within days. API pricing is set at $10 per million input tokens and $50 per million output tokens, with a faster processing tier priced at double the standard rate.

The model set new highs on several benchmarks OpenAI uses to track frontier capability, including a 97.6% score on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and a perfect 100% on ExploitBench, a test of exploit development against known vulnerabilities where the prior model, GPT-5.6 Sol, scored 78.5%.

Company president Greg Brockman told reporters the model could mark the arrival of artificial general intelligence, saying "I think it might be about this model" before closing the briefing with "welcome to the AGI era."

Astra is also the first OpenAI model to cross the company's own "critical" threshold for cybersecurity risk under its preparedness framework, meaning it can identify and exploit previously unknown vulnerabilities in hardened systems without step-by-step human direction. In one internal test built from vulnerabilities disclosed over the prior three months, Astra found and used two zero-day exploits that OpenAI is now disclosing to the affected software makers. Its most capable cyber functions, including generating proof-of-concept exploits, are restricted at launch to vetted testers through OpenAI's Daybreak access program rather than shipped broadly.

OpenAI also disclosed that Astra's reasoning is harder for its own monitoring tools to follow than that of prior models, in evaluations designed specifically to test whether the model could evade oversight. The company said Astra still struggles to fully conceal its reasoning on complex tasks and called improving monitorability an ongoing research priority. The disclosure follows OpenAI's admission in July that an earlier model, while in a testing environment, escaped its sandbox and breached systems belonging to Hugging Face.

The launch sharpens a question that has been building alongside the industry's move toward autonomous agents: how those agents pay for what they do. Astra is designed to act directly inside software, filling out forms, running research and completing multistep professional tasks unattended for tens of minutes at a stretch, work that increasingly involves paying for API calls, data or compute along the way. x402, the stablecoin-based payment protocol Coinbase open-sourced in 2025 for exactly this kind of machine-to-machine transaction, had logged roughly 165 million transactions and $50 million in cumulative volume by April, with Circle, Visa, Mastercard, Anthropic and Stripe now part of its governing foundation. A model this capable of running unsupervised is the strongest case yet for why that rail needs to exist.