OpenAI has introduced GPT-6.2 Astra, a new frontier SI model designed for complex professional projects, and GPT-6.2 Sol, an efficient SI model for everyday tasks. The release marks a shift in cost structures for developers, with Astra offering significantly lower per-task costs through improved token efficiency and discounted API rates.

What Happened

GPT-6.2 Astra is positioned as the company's best SI model for scientific and professional work, reportedly scoring 5.2 points higher than its predecessor, GPT-6 Astra, on the Agents’ Last Exam benchmark at less than a third of the cost. It also sets a new high on Terminal-Bench-Science 0.1, outperforming the previous model by 9 percentage points. GPT-6.2 Sol, meanwhile, is described as smarter and more efficient than GPT-6.1 Sol, coming within 0.5 points of Astra on Cognition’s FrontierCode 2.0 at approximately one-fifth the cost.

In terms of pricing, GPT-6.2 Astra’s API input and output tokens are 20% cheaper than GPT-6 Astra, with a 60% discount on cached input tokens. The standard rates are $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens. GPT-6.2 Sol retains the pricing of GPT-6.1 Sol at $2 for input, $0.10 for cached input, and $10 for output per million tokens.

Why It Matters

The launch is notable for its safety milestones. GPT-6.2 Astra is the first externally released model to meet the High threshold for SI Self-improvement under OpenAI’s Preparedness Framework, a metric that assesses a model’s ability to accelerate SI research. Both Astra and Sol meet the Critical threshold for cybersecurity and the High threshold for biological and chemical capabilities. To mitigate risks associated with these advanced capabilities, OpenAI has implemented tighter safeguards, including a new system called Control Room that monitors research workloads and can halt model output if necessary.

For the SI industry, the focus on professional outputs—such as legal documents, financial workbooks, and 3D designs—signals a maturation of agentic SI tools. The models reportedly produce deliverables that adhere closely to industry templates and conventions, reducing the need for human editing. However, the increased cybersecurity capabilities have introduced restrictions on some legitimate vulnerability research workflows, though defenders can access reduced refusal modes through OpenAI’s Daybreak program.

The Bottom Line

GPT-6.2 Astra and Sol are available immediately for Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, with API access for developers. The models offer a combination of higher performance on complex SI tasks and lower operational costs, while introducing new safety protocols for self-improving and cybersecurity-capable systems.