OpenAI launches GPT-6.2 Astra and Sol models for professional use
The new models offer a split approach between high-performance scientific tasks and efficient daily workflows at lower costs.
The short version
- GPT-6.2 Astra is designed for demanding scientific, legal, and engineering projects.
- GPT-6.2 Sol focuses on managing everyday workflows and routine tasks.
- Astra offers a 20% discount on API rates compared to the previous GPT-6 Astra model.
- Both models meet the Critical threshold for cybersecurity under OpenAI's Preparedness Framework.
- The models are available to Plus, Enterprise, Pro, and Edu subscribers.
Why it matters
This release signals a shift toward specialized AI models that balance high-level performance with lower operational costs. For businesses, it means more affordable access to tools capable of handling complex technical work like scientific research or financial auditing.
A split approach to AI tasks
OpenAI has officially introduced two new models, GPT-6.2 Astra and GPT-6.2 Sol, to serve different types of users. The company designed these models to address a wide variety of professional and routine tasks by separating high-level complexity from everyday efficiency. These tools are now available to various subscribers, including Plus, Enterprise, Pro, and Edu users within ChatGPT Work and Codex.
The distinction between the two models is based on the nature of the work being performed. GPT-6.2 Astra is positioned as a high-performance option intended for demanding projects in scientific and professional domains. In contrast, GPT-6.2 Sol is built for the efficient management of everyday workflows, making it a more practical choice for routine operations.
Developers can access these new models through the OpenAI API. Additionally, the company is offering Ultrafast access for Pro 500 and Pro 200 subscribers. This rollout marks the first time that Pro 200 subscribers will have access to the Ultrafast feature, starting with the GPT-6.2 Sol model.
Performance in technical fields
GPT-6.2 Astra is specifically tailored for technical applications such as engineering design, legal work, and finance. According to OpenAI, the model is intended to handle complex projects while maintaining a lower cost than previous versions. It reportedly achieved superior results in various industry benchmarks, including tests for biological data interpretation and experiment simulation.
In the Agents’ Last Exam, which evaluates multi-step professional tasks, GPT-6.2 Astra outperformed GPT-6 Astra by 5.2 points. It also beat Claude Fable by 5.5 points, which the company claims happened at a fraction of the associated costs. These results suggest the model is becoming more capable at handling multi-step logic without requiring as much computing power.
The model also showed strength in scientific environments. In the Terminal-Bench-Science 0.1 test, which focuses on scientific tasks within a terminal environment, GPT-6.2 Astra scored nine percentage points higher than the previous GPT-6 Astra. These improvements are particularly relevant for users who need to process complex data in specialized technical settings.
Cost-efficiency and accuracy
One of the primary goals for the new models is to improve cost-efficiency and token usage for developers. For GPT-6.2 Astra, the API rates are set at $8 per million input tokens and $40 for output tokens. This represents a 20% discount compared to the previous GPT-6 Astra model, while cached inputs see a 60% reduction.
The GPT-6.2 Sol model keeps the same pricing structure as the GPT-6.1 Sol. It costs $2 per million for input, $0.10 for cached input, and $10 for output. By offering these different pricing tiers, OpenAI allows users to choose the most economical model for their specific needs, whether they require high-end reasoning or high-volume routine processing.
Beyond cost, OpenAI highlights significant improvements in accuracy for specific industries. In the legal, finance, and consulting fields, the models are designed to adhere more closely to professional templates. For example, during tests by Artificial Analysis, GPT-6.2 Astra correctly handled 62.5% of questions over five attempts when analyzing complicated financial reports.
Safety and risk management
OpenAI has implemented specific safeguards for these models to manage risks associated with high-stakes tasks. Both GPT-6.2 Astra and GPT-6.2 Sol reach the Critical level for cybersecurity within the company's Preparedness Framework. This framework helps the company monitor and limit the risks associated with tasks that could be used for malicious purposes.
The company added new training signals to improve the defensive capabilities of the models. OpenAI noted that developers often rely on coding agents to write software, which can inadvertently lower the barrier for finding and exploiting vulnerabilities. To counter this, the company has strengthened controls around the training and deployment of these specific models.
The models also reach the High level for biological and chemical abilities, as well as for AI self-improvement, as defined by the company's internal framework. OpenAI has designed safeguards to monitor outputs and restrict dual-use cases that involve biological or cybersecurity concerns. These measures are intended to ensure that the models remain useful for professional work while minimizing potential harms.
The regulatory landscape
The release of these advanced models comes at a time of increased scrutiny for the company. Recently, California Attorney General Rob Bonta filed a lawsuit against OpenAI regarding its safety protocols. The lawsuit claims that the company did not implement adequate oversight during the development and testing of its advanced AI models.
In response to these concerns, OpenAI continues to emphasize its Preparedness Framework and the specific safeguards built into the GPT-6.2 series. The company is balancing the need to provide powerful tools for scientific research and engineering with the necessity of preventing misuse. These safeguards are a central part of how the company manages the deployment of its most capable technology.
Moving forward, the success of these models will likely depend on how well they perform in real-world professional environments. Users in finance, law, and science will be the first to see if the promised accuracy and cost-savings hold up. The company will need to maintain these safety standards as the models become more integrated into production software and research workflows.
What to watch
- How the 20% cost reduction for Astra affects the adoption of AI in small-scale scientific research.
- Whether the new cybersecurity safeguards successfully prevent the exploitation of vulnerabilities in production software.
- The outcome of the California Attorney General's lawsuit regarding OpenAI's safety protocols.
Try it
If you are a Plus or Pro subscriber, check your ChatGPT Work or Codex dashboard to see if you have access to the new GPT-6.2 Astra or Sol models.
Sources and links
First reported by techmonitor.ai. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.