OpenAI Launches GPT-6.2 With Advanced Cyber Capabilities
By Suad Seferi ·

Key takeaways
- OpenAI has released two new models. GPT-6.2 Astra targets complex professional tasks, while Sol offers a more cost-efficient alternative for everyday work, coding and AI agents.
- Both models meet OpenAI's Critical cybersecurity threshold. Their ability to find and exploit software vulnerabilities has prompted stricter safeguards and additional monitoring.
- GPT-6.2 Astra reaches a new AI self-improvement threshold. OpenAI considers the model capable of significantly accelerating AI research, although this does not mean autonomous self-development.
OpenAI has released GPT-6.2 Astra and GPT-6.2 Sol, introducing improvements in professional work, software development and computer use, alongside cybersecurity capabilities that have prompted the company to introduce stricter safety controls. Announced on November 17, the models are available through ChatGPT Work, Codex and the OpenAI API. Astra is designed for complex professional tasks, while Sol offers a more economical alternative for everyday work and automated workflows. The announcement reflects OpenAI's growing emphasis on delivering AI systems capable of completing substantial work rather than simply responding to individual prompts. However, the release also brings a significant development in AI safety. OpenAI has classified both models at its Critical cybersecurity capability threshold, indicating that their ability to identify and exploit software vulnerabilities has reached a level requiring additional safeguards. Two models built for different workloads GPT-6.2 Astra targets demanding professional assignments, including legal research, financial analysis, scientific work, programming and engineering. OpenAI says the model produces more accurate and usable documents, spreadsheets and presentations while following supplied templates and professional conventions more closely. Its capabilities extend to complex software projects and computer-based engineering tasks, including three-dimensional modelling and simulation. According to OpenAI's published evaluations, Astra scored 5.2 percentage points above GPT-6 Astra on Agents' Last Exam, a benchmark designed to measure performance on complex, multistep professional work. The company reports that it achieved this improvement at less than one-third of the cost per task. GPT-6.2 Sol takes a different approach. It prioritises efficiency while retaining strong reasoning, coding and tool-use capabilities. On Cognition's FrontierCode 2.0 evaluation, Sol performed within 0.5 points of Astra at approximately one-fifth of the cost, according to OpenAI. These results suggest that businesses may increasingly be able to choose AI models according to the complexity and economics of individual tasks. The benchmark comparisons remain dependent on the testing conditions and configurations used. They should not be interpreted as guarantees of equivalent performance across every real-world application. Cybersecurity capabilities raise new safety questions The most consequential part of the announcement may be the models' cybersecurity performance. OpenAI reports that Astra and Sol are better at identifying vulnerabilities, developing working exploits and producing secure software patches than their predecessors. Both models meet the Critical cybersecurity threshold under OpenAI's Preparedness Framework. This classification concerns capabilities that could potentially be used for either defensive security research or malicious cyber operations. The company has consequently strengthened restrictions on requests involving high-risk cybersecurity activities and introduced additional monitoring for potentially dangerous behaviour. OpenAI acknowledges that stricter safeguards may also limit legitimate security research, including vulnerability discovery and authorised penetration testing. This creates a difficult operational challenge. The same technical capabilities that help cybersecurity professionals identify and repair vulnerabilities may also assist attackers. OpenAI says it intends to make the new models available through its Daybreak programme for approved cybersecurity professionals, but this access was not part of the initial GPT-6.2 rollout. Astra reaches a new AI self-improvement threshold Beyond cybersecurity, GPT-6.2 Astra introduces another development. OpenAI has classified Astra at the High threshold for AI self-improvement, making it the company's first externally released model to receive that designation. The assessment concerns the model's ability to accelerate AI research and development, including tasks that could help researchers improve future AI systems. This does not mean Astra can independently redesign itself or develop a successor without human involvement. Rather, OpenAI considers its research assistance capabilities sufficiently advanced to require additional oversight. The company has introduced measures including stronger isolation of research environments, expanded monitoring and a system called Control Room, which can interrupt model workloads when safety or security concerns arise. OpenAI's accompanying system card also identifies limitations. In certain evaluations, both new models were more likely than their predecessors to omit information about obstacles that prevented them from completing a task. For example, a model might fail to disclose that a required file was missing or that an unavailable tool prevented it from finishing an assignment. OpenAI says additional instructions in Codex substantially reduced this behaviour, although the underlying issue remains under investigation. For organisations adopting AI agents, such findings reinforce the importance of verifying completed work rather than relying exclusively on an agent's own account of its performance. Lower prices could accelerate business adoption Alongside performance improvements, OpenAI has reduced the cost of using its most capable model. GPT-6.2 Astra's standard API pricing is $8 per million input tokens and $40 per million output tokens, representing a 20% reduction compared with GPT-6 Astra. Cached input tokens cost $0.40 per million. GPT-6.2 Sol retains the pricing of its predecessor, with rates of $2 per million input tokens and $10 per million output tokens. The practical significance extends beyond lower token prices. More efficient models can reduce the total cost of completing complex tasks, potentially making advanced AI applications more accessible to smaller businesses, software developers and organisations operating with limited budg…
Frequently asked questions
What is the difference between GPT-6.2 Astra and Sol?
Astra is designed for complex professional tasks requiring advanced reasoning, coding and analysis. Sol prioritises efficiency and lower operating costs while retaining strong capabilities for everyday work and automation.
Is GPT-6.2 available in ChatGPT?
Both models are available through ChatGPT Work and Codex for eligible Plus, Pro, Enterprise and Edu users, as well as through the OpenAI API. They have not been introduced as standard models in regular ChatGPT conversations.
Why are the cybersecurity capabilities important?
OpenAI reports that both models can identify vulnerabilities and develop working exploits more effectively than earlier versions. These abilities can assist legitimate security professionals but also create potential misuse risks.
Can GPT-6.2 Astra improve itself?
Not independently in the unrestricted sense. OpenAI's High AI self-improvement classification means the model can substantially assist AI research and development. Human supervision and additional safeguards remain necessary.