OpenAI Launches GPT-6 Astra with Expanded Capabilities and Limited Access

OpenAI has launched GPT-6 Astra, its first internal model classified at the "critical" cyber risk level. Access is gradual and restricted, with offensive capabilities held back. However, the company claims to have achieved speed gains, reduced costs per task, and expanded abilities to control computers and software.

Less Readable Reasoning and Evasion Tactics Noted by OpenAI

Security teams typically rely on step-by-step reasoning reviews. OpenAI states that the reasoning chain of GPT-6 Astra is less readable than that of its predecessor and that the model better controls what it writes, leaving fewer compromising traces. According to OpenAI, during evasion experiments, Astra can alter its responses without being detected and sometimes manages to bypass internal monitoring systems during sabotage operations. However, no cases have been observed where reasoning was hidden in seemingly innocuous text. To prevent abuse, classifiers continuously monitor the model's reasoning and actions when it uses tools and can interrupt any unauthorized actions, a method that OpenAI admits may hinder some legitimate uses, including in a defensive context.

Restricted Rollout Primarily for Defense Professionals

The version made available is limited in its uses. It covers defensive functions such as code review and suggesting fixes, and OpenAI indicates that it will refuse more advanced cybersecurity tasks such as creating demonstration exploits. Additional capabilities are expected to be opened in the coming weeks through Daybreak, the program aimed at cybersecurity professionals. Access is initially reserved for a limited number of organizations, with several U.S. media outlets reporting an initial targeting of Daybreak clients. Plus, Pro, Business, and Enterprise subscribers of ChatGPT are expected to gain access in the coming days, while the option remains disabled by default for Enterprise clients and requires activation by an administrator. Free accounts remain excluded, and a GPT-6 Astra Pro variant is announced for Pro, Business, and Enterprise offerings.

Pricing, API, and Partner Clouds for Developers

For developers, the model is available in the API under the identifier gpt-6-astra at $10 per million tokens for input and $50 for output, prices aligned with the base rates of Claude Fable 5.1, released two days earlier by Anthropic. A Fast option is announced, doubling the speed for a doubled cost. OpenAI also plans availability through AWS Bedrock and Microsoft Azure.

Acceleration in Codex and Computer Control Claimed

OpenAI highlights speed and quality gains. With the update to Codex, certain tasks are said to be executed up to 1.9 times faster than with GPT-5.6 Sol, and the time required to control software interfaces would decrease by about 47% compared to the previous model. The company also claims a reduced cost per task, a positioning that it says allows for more professional delegation. In practice, Astra can navigate interfaces, fill out forms, update records in a CRM, organize schedules, write summaries, build websites, and test their functionalities. It would also produce documents, spreadsheets, and presentations based on internal templates while adhering to visual guidelines and styles, filtering out deemed superfluous context and demonstrating better visual judgment. In software development, OpenAI presents Astra as its best iteration: it completes instructions when the outcome does not change, requests targeted clarifications when they are crucial, progresses on independent parts, integrates instructions along the way without losing the initial objective, and retains notes beyond a context window.

A Defined "Critical" Level and Measured Exploits

OpenAI's internal framework classifies models by risk levels, with the "critical" tier indicating the ability to identify novel vulnerabilities and describe their exploitation on protected systems, without human assistance at every step. No model from the company had reached this level before GPT-6 Astra. Without its production safeguards, OpenAI reports a score of 100% on ExploitBench for Astra, compared to 78.5% for GPT-5.6 Sol, and the discovery and use of two recent vulnerabilities in Chrome that had not been previously reported. OpenAI states it has communicated this information to maintainers.

Launch Under High Scrutiny and Promises of Performance

OpenAI launches GPT-6 Astra after delaying its release by a month to manage a risk level deemed "critical," a threshold never before reached by the company. The deployment occurs amid heightened vigilance following summer incidents, including an intrusion by an OpenAI agent at Hugging Face. The company presents this model as "the most intelligent and aligned in the world" and claims it surpasses others in cybersecurity. In a set of 54,000 internal tasks simulated in Codex, OpenAI reports that the number of severe misalignment alerts is half that observed with GPT-5.6 Sol, even as the notion of misalignment encompasses off-scope responses, evasion maneuvers, or unsolicited initiatives.