What Happened?
OpenAI has officially launched GPT-6 Astra, a frontier AI model the company describes as “the world’s most intelligent and aligned model”. The announcement came during a closed press briefing where OpenAI President Greg Brockman made an unusually direct claim: “For me personally, I do think we’re there,” he said, before concluding, “Welcome to the AGI era.”
Astra represents a generational leap from its predecessor, GPT-5.6 Sol, which launched just months earlier in July 2026. The model was trained on more than 100,000 GPUs at OpenAI’s Stargate facility in Texas—the company’s largest training run to date. It features a 1.05-million-token context window and can generate up to 128,000 output tokens, enabling it to handle large, complicated workflows.
Perhaps most notably, Astra is the first OpenAI model to reach the “Critical” cybersecurity capability level under the company’s Preparedness Framework—a designation that triggered additional deployment restrictions and safeguards. The model autonomously discovered two previously unknown zero-day vulnerabilities during internal evaluations, which were disclosed to maintainers.
Key Features & Technical Breakdown
GPT-6 Astra’s benchmark performance is nothing short of extraordinary. The model achieved a 98% score on FrontierMath Tier 4 (advanced mathematical reasoning), 99.9% on ARC-AGI-3 (abstract reasoning), and a perfect 100% on ExploitBench (cybersecurity exploit development). On OSWorld 2.0, which tests computer-use capabilities, Astra scored 72.6% and completed tasks in about 40 minutes, compared with 65.7% and roughly 75 minutes for GPT-5.6 Sol.
The model’s computer-use capabilities represent a fundamental architectural shift. Rather than requiring developers to build API integrations for every application, Astra is designed to navigate software the way a human does—working across browsers, spreadsheets, websites, and desktop applications. It can fill out online forms, update CRM records, organize calendars, conduct web research, analyze scientific data in Python notebooks, create and test websites, and even operate engineering applications like KiCad and FreeCAD.
API access is priced at $10 per million input tokens and $50 per million output tokens—approximately 2.5 times the cost of GPT-5.6 Sol. Enterprise administrators must manually enable Astra for their workspaces, with access turned off by default at launch. The model’s knowledge cutoff is April 30, 2026, and it supports tools including web search, file search, computer use, code execution, and image generation.
Industry Impact & Market Reaction
The AGI declaration has already sparked intense debate. Nvidia CEO Jensen Huang took to X to declare that “AGI has arrived” following Astra’s launch. However, AI experts have questioned the validity of Brockman’s claim, noting that OpenAI’s official documentation focuses on capabilities and safety measures rather than formally declaring AGI.
Competitive pressure is mounting. OpenAI is positioning Astra directly against Anthropic’s Claude Fable 5.1 (which scored 55.8% on Terminal-Bench 4.0 versus Astra’s 57.9%) and Google’s Gemini. The race has shifted decisively from which model produces the strongest response to which can reliably take action, navigate software, and complete meaningful work with less human intervention.
On launch day, ChatGPT, Claude, and Grok all experienced simultaneous outages affecting conversations, logins, file uploads, voice mode, and image generation. The exact cause remains unclear, with no confirmed evidence that Astra caused the disruption.
For enterprises, the implications are profound. Analyst Sanchit Vir Gogia of Greyhound Research noted that Astra’s “Critical” label is a disclosure event rather than a capability event, arguing that “Astra is now the only frontier model whose cyber capability an enterprise actually knows” because it’s the only one measured against a published threshold. OpenAI is initially restricting the model’s most advanced cybersecurity capabilities to trusted defenders through the Daybreak Blue programme.
