On September 3, OpenAI officially unveiled its new flagship model — GPT-6 Astra. The release marks not just an evolution of computing power, but a qualitative leap toward full autonomy for AI agents. At the initial stage, access to the model is limited to the Trusted Access program for select organizations, but in the coming days it will become available to ChatGPT Plus, Pro, Business, and Enterprise users, as well as via the API and Amazon Bedrock. In parallel, Microsoft is integrating Astra into its Microsoft Foundry platform, expanding access for Azure enterprise customers.
Agentic intelligence on a new level
The key difference between Astra and its predecessor, GPT-5.6 Sol, is its ability to independently interact with the computer environment. The model can work with browsers and applications: filling out forms, updating CRM records, analyzing data, and even testing software. Test results are impressive: on a subset of OSWorld 2.0, Astra scored 72.6% versus 65.7% for Sol, and it completes tasks in latency simulation 47% faster (~40 minutes versus 75). In the Agents' Last Exam benchmark, the result was 59.3%, higher than both GPT-5.6 Sol (53.6%) and Claude Opus 5 (55.5%).
Progress in the scientific field deserves special attention. In the FrontierMath Tier 4 test, the model achieved an impressive 97.6%. Moreover, developers claim that Astra helped improve two results in number theory, one of which had remained unchanged for over 80 years. This is the first time an AI has made such a significant contribution to fundamental mathematics, opening new horizons for scientific research.
Critical-level cyber capabilities
However, the main topic of discussion is safety. Astra became the first OpenAI model to reach the "Critical" level of cyber capabilities under the internal Preparedness Framework system. In ExploitBench tests, it showed 100% without production restrictions, and in expert trials it was able to create exploits for zero-day vulnerabilities. This forced the company to implement strict restrictions in the public version: the model will refuse advanced cyber tasks, and access for defensive scenarios will be provided through the new Daybreak program.
The company acknowledges that monitoring Astra's written reasoning has become more difficult, requiring additional control measures. This raises a fundamental question: are we ready for the emergence of an AI that can both defend and attack with equal effectiveness?
Investments in defense and commercialization
Alongside the release, OpenAI announced the Daybreak for Frontline Defenders program worth $1 billion. The initiative aims to subsidize access to AI tools for critical infrastructure operators, banks, and open-source software developers in the United States. This is a timely step, given that the cyber arms race is entering a new phase.
The cost of using the model also raises questions. At $10 per 1 million input tokens and $50 for output tokens, along with multiplier factors for long queries, highly complex analytical tasks could prove to be an expensive endeavor. The context window of 1.05 million tokens is an impressive figure, but it requires significant computing resources.
My expert assessment: GPT-6 Astra is not just another update, but a transition to a new paradigm where AI becomes a full-fledged digital employee. However, reaching the "critical" level of cyber capabilities is a double-edged sword. The market will need to closely monitor how OpenAI balances innovation and safety, and whether this technology will become a new battlefield for malicious actors.