OpenAI has introduced two models at once: GPT-6.2 Astra and GPT-6.2 Sol. Their input and output prices are $8 and $40 per million Tokens—the basic unit used to measure text processed by a model—and $2 and $10, respectively. What we think deserves attention is not the version number, but the lower cost of invoking these models for complex tasks, even though industry evaluations do not show a comprehensive lead.

What this is

Astra targets difficult projects, with a focus on capability and complex tasks; Sol is designed for everyday work, emphasizing efficiency and versatility. Astra is 20% cheaper than its predecessor, while its cached-input price has fallen from $1 to $0.40 per million Tokens. Caching allows already processed content to be reused across repeated calls, making it especially valuable for workflows that run continuously or repeatedly read the same material. Sol, meanwhile, retains its predecessor’s price.

Developer Simon Willison has backed both new models and compared them on an animation-generation task. In his tests, however, neither model outperformed Claude Fable 5.2 or Gemini 4 Carbon, showing that lower prices do not necessarily translate into superior capabilities.

Industry view

Our view is that GPT-6.2’s commercial value comes first from cost optimization, not a breakthrough in benchmark scores. Astra’s price reduction is especially significant for enterprise workflows involving multiple calls and multistep execution. Sol’s unchanged pricing, however, suggests that it fills a gap in the product lineup rather than representing a broad price cut.

The risks are equally clear. Astra did not beat Claude Fable 5.2 on some benchmarks, and developers have reported that its way of expressing itself can be confusing. A more practical problem is that long-running tasks may invoke multiple submodels to work together. If the system suddenly asks users about internal details they cannot understand or intervene in, the process may become unmanageable even if the final result is usable. The longer a model can work, the more important it becomes to know whether errors will be amplified and whether its process remains inspectable.

Impact on regular people

For enterprise IT: Procurement teams should not compare only the strongest model. They should also tier models by task difficulty, assigning high-frequency routine work to lower-cost models and reserving higher-priced models for complex projects. This could reduce overall budgets.

For individual professionals: Most people are more likely to use these models indirectly through productivity software than to choose an API version themselves. Whether lower prices translate into more stable and useful products will still depend on how well each tool executes tasks.

For the consumer market: Model capabilities have not diverged enough to create an exclusive advantage, so developers may test several platforms at once. Consumers are most likely to feel lower prices for comparable AI features, as well as a shift in how products allocate computing resources—from simply demonstrating technical prowess toward completing tasks continuously.