OpenAI released GPT-6 Astra on September 3, 2026, billing it as “the world’s most intelligent and aligned model” — and company president Greg Brockman went a step further. Asked whether the new model marks the arrival of artificial general intelligence, he answered: “For me personally, I do think we’re there.” OpenAI itself stopped short of an official AGI declaration; Brockman noted that AGI is no longer a contractual trigger for the company and has become more of a “mission concept or spiritual concept.”
The launch numbers are eye-catching. Astra posted 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and a perfect 100% on ExploitBench, and it solved SRE-Bench binary reverse-engineering tasks with 99.2% success within four attempts. Long-context recall is just as striking: 100% on an eight-needle retrieval test at 256K–512K tokens and 96.3% between 512K and 1M tokens. Independent scoring is more sober — Artificial Analysis put its Intelligence Index at 61, tied with GPT-5.6 Sol — but on agentic work the story improves: Astra’s Coding Agent Index landed 2 points above Sol at less than half the cost of Claude Fable 5.
Computer use is the marquee skill. OpenAI says Astra outperforms rivals at driving a browser and desktop software, and in one demonstration the model autonomously used Google Maps to find a pediatrician, visited the practice’s website and filled in the contact details on its own.
The same strengths cut both ways. Astra is the first OpenAI model to reach the “Critical” level for cybersecurity capabilities under the company’s Preparedness Framework: given the right tools and access, it can find previously unknown security flaws and develop ways to exploit well-protected systems without a human guiding every step. OpenAI responded with guardrails — the deployed versions block access to the most advanced cyber capabilities, tool-based sessions get an extra layer of monitoring, and the model is more resistant to jailbreak attempts than GPT-5.6 Sol. Less reassuring: Astra relies on a reasoning technique called opaque recurrence, which makes its chain of thought harder for researchers to follow, and OpenAI acknowledged that the model’s monitorability has decreased relative to Sol.
At release on September 3, 2026, Astra was live for a limited set of organizations, with ChatGPT Plus, Pro, Business and Enterprise subscribers, the API and Amazon Web Services slated to follow in the days after launch. In the API the model is labeled gpt-6-astra and costs US$10 per million input tokens and US$50 per million output tokens.
Earlier on GismoLand: OpenAI Astra. When an AI gets too good at breaking into things.
