OpenAI Launches GPT-6 Astra, Its First Model Rated 'Critical' for Cyber Risk
News Summary
OpenAI officially launched GPT-6 Astra on September 3, 2026, describing it as a "generational leap" over its predecessor and the first model the company is willing to associate with the arrival of artificial general intelligence (AGI). President Greg Brockman said Astra represents a step change in reasoning, computer use, and software engineering capability, while OpenAI simultaneously classified the model as "Critical" under its internal Preparedness Framework for cybersecurity risk — a first for any OpenAI release.
Rollout and Availability
Astra is rolling out in stages rather than all at once. The first group to receive access is made up of participants in OpenAI's Daybreak program, an application-based initiative for cybersecurity defenders, effective Thursday, September 3, 2026 (Pacific Time). Broader availability for ChatGPT Plus, Pro, Business, and Enterprise subscribers, as well as developers using the OpenAI API, is expected "in the coming days," according to the company. The model is also being made available through Amazon Bedrock, and API access is offered under the identifier gpt-6-astra, with a corresponding Astra Pro tier for higher-usage subscription plans.
API pricing for developers is set at $10 per million input tokens and $50 per million output tokens in standard mode, roughly 2.5 times the cost of the company's prior flagship model, GPT-5.6 Sol. A faster inference mode, promising about 2.5x lower latency on comparable tasks, is priced at double the standard rate.
Capabilities and Benchmarks
OpenAI positioned Astra as state-of-the-art across computer use, web browsing, software engineering, cybersecurity, and scientific reasoning tasks. Brockman highlighted computer use as a headline feature, saying the model can "zip through spreadsheets, fill out forms, and navigate across web pages" at speeds he described as superhuman for routine office tasks.
On published benchmarks, Astra scored 72.6% on OSWorld 2.0, a computer-use evaluation, while completing tasks in roughly 47% less time than GPT-5.6 Sol. The company also reported that Astra saturated Tier 4 of FrontierMath, a demanding mathematics benchmark, with a 97.6% score, and reached 100% on ExploitBench, an internal evaluation of vulnerability discovery and exploitation skill that fed directly into the model's "Critical" safety classification.
Cybersecurity Safeguards and the Critical Classification
The same capabilities that make Astra useful for defenders also raised internal alarms at OpenAI. The company said Astra is the first model able to autonomously identify previously unknown security vulnerabilities and construct working exploits against well-protected systems without step-by-step human direction, prompting its designation as "Critical" under the Preparedness Framework, OpenAI's internal policy for scaling safety precautions to a model's risk level.
To manage that risk, OpenAI said it delayed portions of Astra's development over several weeks to add and test additional safeguards before proceeding with the launch. Access to the model's most advanced cybersecurity abilities is being restricted to a limited set of vetted testers under the Daybreak program, and even within that program, approved users are permitted to use Astra for common cyber-defense tasks but are blocked from using it to develop exploits or plan offensive cyberattacks. Wider defensive access is planned through a further tier called Daybreak Blue as testing continues.
Industry Reaction
Coverage of the launch noted that Astra's benchmark results place it ahead of rival frontier models, including comparisons to Anthropic's Fable 5.1 on several published leaderboards, intensifying competition among leading AI labs on reasoning, coding, and agentic computer-use tasks. Brockman's suggestion that Astra may already meet or be within reach of OpenAI's internal definition of AGI — an AI system that outperforms humans at most economically valuable work — drew significant attention from industry observers, though OpenAI stopped short of declaring the threshold formally met.
What Happens Next
OpenAI has not published an exact date for when general ChatGPT and API access will open beyond the initial Daybreak cohort, saying only that the expansion will happen within days of the September 3, 2026 announcement (Pacific Time). The company said it will continue to monitor real-world usage under the Preparedness Framework and may adjust safeguards, particularly around cybersecurity-related features, as the rollout widens to enterprise and consumer audiences.