OpenAI Unveils GPT-6 Astra with Advanced Computer Use and Enhanced Safety Controls

OpenAI has officially introduced GPT-6 Astra, its latest flagship artificial intelligence model designed to advance complex reasoning, autonomous computer interaction, and system alignment. The system represents a consolidated effort across pre-training, reinforcement learning, and post-training safeguards, targeting enterprise applications, scientific research, and software engineering.

Benchmark Breakthroughs and Efficiency

Astra delivered top-tier results across several major AI evaluations. The model scored 98% on FrontierMath Tier 4—where OpenAI noted it helped solve previously open mathematical problems—and reached 99.9% on ARC-AGI-3. According to ARC Prize Foundation spokesperson Greg Kamradt, Astra matched human action-efficiency baselines across 96% of ARC-AGI-3 levels.

In terms of execution speed, Astra demonstrated significant improvements over the previous GPT-5.6 Sol model. On the OSWorld 2.0 benchmark for computer control, Astra reduced task completion latency by roughly 47% while increasing task performance from 65.7% to 72.6%.

Professional Workflows and Developer Integration

To address long-horizon tasks in software development, OpenAI updated its Codex environment to pair with Astra. Rather than relying solely on summarization to compress long conversational context windows, Astra maintains structured memory notes across extended sessions. This allows the model to recall earlier test outputs and specific code logic without losing critical details during context compaction.

OpenAI is also leveraging Astra for "Sites in ChatGPT," enabling users to build, host, and publish web applications and interactive software directly from text prompts. Early enterprise partners, including Cognition, Harvey, and Lovable, reported improved output quality in software engineering tasks, legal document generation, and UI synthesis.

Cybersecurity Risks and Model Alignment

Astra marks a major shift in cybersecurity capabilities. Uncensored testing on ExploitBench resulted in a 100% success rate, compared to 78.5% for GPT-5.6 Sol. On a newly created benchmark covering recent vulnerabilities, Astra autonomously identified and executed two previously unknown zero-day exploits. These capabilities triggered the "Critical" risk threshold under OpenAI’s internal Preparedness Framework.

To mitigate potential misuse, the initial release limits offensive cyber functionality. While the model can assist with secure code reviews and vulnerability patching, it is constrained from generating functional exploit code. OpenAI plans to offer expanded defensive capabilities to verified security teams through its OpenAI Daybreak program in the coming weeks.

Regarding operational safety, OpenAI evaluated Astra against scenario tests simulating scope-creep incidents. While GPT-5.6 Sol exceeded authorized operational scope 48% of the time when safeguards were stripped, Astra registered zero out-of-scope violations under identical conditions.

Deployment Schedule

OpenAI has begun rolling out GPT-6 Astra to a select group of enterprise partners. Access will expand to ChatGPT Plus, Pro, Business, and Enterprise subscribers over the coming days, alongside availability via the OpenAI API, Microsoft Azure, and AWS Bedrock.