OpenAI has introduced GPT-6.2 Astra and GPT-6.2 Sol, two models designed to make increasingly capable AI work more economical. Announced on November 17, the release puts less emphasis on a single headline intelligence score and more on how much useful work a model can complete for a given budget.
Astra is positioned for difficult professional assignments, including software engineering, scientific analysis and work with documents, spreadsheets and presentations. Sol is pitched as a faster, less expensive everyday model. The company describes both as part of its new model lineup, according to OpenAI's launch announcement.
Lower prices are only part of the story
For API customers, GPT-6.2 Astra is priced at $8 per million input tokens, $40 per million output tokens and $0.40 per million cached input tokens. OpenAI says the ordinary input and output rates are 20% below those of GPT-6 Astra. OpenAI also presents Sol as a lower-cost option for routine work, although customers should confirm the pricing for the product and service tier they actually use.
That price list does not automatically reveal the cost of finishing a real assignment. A model that produces fewer unnecessary tokens, needs fewer retries or makes fewer coding mistakes can be cheaper per completed task even with a higher per-token rate. Conversely, difficult workloads can still become expensive through extended reasoning, tool calls and human verification.
OpenAI says Astra outperforms its predecessor on long-running professional tasks and science evaluations, while Sol narrows the coding-quality gap at a lower cost. These are the company's reported results, obtained using specified benchmark setups, not an independent guarantee that every enterprise will see equivalent gains. The launch announcement provides benchmark descriptions, but organizations should check the details of each test before treating the comparisons as purchasing evidence.
Why the release matters for businesses
Organizations increasingly have to decide which AI system should handle each step in a workflow. A routine support summary may not justify a high-priced reasoning model. Reviewing a complex code change or examining experimental data may. The practical promise of a two-model lineup is the ability to move between those tiers without redesigning an entire application.
That is particularly relevant as businesses attempt to put agents into production. An agent's total cost includes its context window, repeated interactions with external systems, verification steps and the effort needed to recover from errors. Claims about low cost per task deserve testing on actual company workflows, not just model leaderboards.
The company's technical claims span professional workflows and AI research, but they do not substitute for independent deployment tests. In its announcement, the company says Astra reaches a high threshold for AI self-improvement capabilities. This is a reported safety-evaluation classification, not proof that deployed systems can autonomously improve themselves without limits.
What remains unproven
The new models arrive amid intense competition to make AI both stronger and more affordable. Microsoft's separate November 17 model announcements underline how quickly suppliers are changing the price-performance equation. Buyers should expect vendor comparisons to use different datasets, model settings and accounting assumptions.
The most useful next step is a controlled evaluation: run representative tasks, measure completed outcomes and error recovery, calculate full API and review costs, and compare against existing models. Those results will say more about GPT-6.2's business value than a universal claim of superiority.