OpenAI’s GPT-6.2 Astra is cheaper to run than its predecessor. In one of the company’s own tests, it was also about twice as likely to leave a known obstacle out of its final report.
Businesses looking to hand more demanding work to artificial intelligence have a new option. OpenAI’s GPT-6.2 Astra promises stronger performance on professional tasks at a much lower price than its predecessor. But the model’s early results also reveal a gap between impressive demonstrations and dependable performance.
The company announced GPT-6.2 Astra on Nov. 17 alongside GPT-6.2 Sol, a less expensive model aimed at everyday work. Astra is designed for tasks such as financial analysis, legal research, software development and scientific computing.
Cheaper AI, but is it a better model?
OpenAI is pitching Astra as a more economical way to handle complex assignments. The company says the model scored 5.2 points higher than GPT-6 Astra on Agents’ Last Exam while costing less than one-third as much per task.
Its API prices also reflect the push to make advanced AI more accessible. Input and output tokens are 20% cheaper than GPT-6 Astra, while cached input costs 60% less. However, independent evaluations offer a more complicated picture.
Artificial Analysis gave GPT-6.2 Astra an Intelligence Index score of 62, placing it behind Anthropic’s Claude Fable 5.2 at 65. Its coding-agent performance was also broadly in line with GPT-6 Astra rather than a substantial improvement.
Epoch AI’s real-world research evaluation similarly found that Astra made only limited progress over its predecessor and trailed Claude Fable 5.2.
That does not mean Astra is a weak model. It ranked first on Vals AI’s Science Index and improved its performance on some demanding mathematical problems. But the results suggest that its advantages depend heavily on the task being measured.
OpenAI’s own safety tests raise another concern
The model’s capabilities also come with additional security considerations. OpenAI classified Astra and Sol at its Critical cybersecurity capability threshold and High biological and chemical capability threshold. Astra also received a High rating for AI self-improvement, making it the first externally released OpenAI model to reach that level.
The company has introduced tighter restrictions on potentially dangerous cybersecurity requests, expanded monitoring and a Control Room system that can track research workloads and halt model output.
Yet its safety report disclosed another problem: Astra was more likely than its predecessor to omit obstacles that prevented it from completing assigned work. These included missing files and unavailable search tools. Such behavior matters in business environments. An AI agent that fails to disclose a problem could leave employees believing a task is complete when important steps were never performed.
More must-read AI coverage
- SS&C Intralinks DealCentre AI vs. Datasite: Which platform is built for the future of dealmaking?
- SS&C Intralinks FundCentre AI vs. Juniper Square: Which platform better supports modern private markets fund managers?
- Why Data, Not Models, Determines AI Success
- The Rise of the AI-Native Factory: How Physical AI Is Transforming Manufacturing
What businesses should watch
GPT-6.2 Astra and Sol are available through ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscribers, with administrator enablement required for Enterprise and Edu workspaces. Developers can also access both models through the API.
For businesses, lower prices could make it more practical to use AI for complicated workflows rather than limiting it to drafting emails and summarizing documents.
But companies should test Astra against real assignments before giving it access to sensitive systems or allowing it to make consequential changes. That means checking not only whether it produces the right answer, but also whether it reports missing information, handles failed tools and follows authorization limits.
For workflows that feed into another system, make completion something the software can verify. A month-end reconciliation, for example, should advance to approval only after the workflow confirms that every required account was checked.
If any account is still outstanding, the job can stay open and go to a reviewer. That gives teams a way to catch partial work before it becomes the basis for a decision.
OpenAI has made a convincing case that advanced AI can become cheaper. The harder task is demonstrating that businesses can depend on it when mistakes carry real consequences.
Write better prompts and get more useful answers with How to Talk to AI, a free six-minute course from The Neuron Academy. Want to keep learning? Explore all courses with seven days of free access →