Frontier Models
GPT-6.2 Astra and Sol turn model choice into a cost decision
Two models released the same day, priced four times apart, performing within half a point of each other on coding. The interesting question is no longer which is best.
In brief
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November across ChatGPT Work, Codex and the API. Astra targets demanding professional work at $8/$40 per million tokens; Sol handles everyday multistep tasks at $2/$10. OpenAI reports Sol lands within half a point of Astra on FrontierCode 2.0 at roughly a fifth of the cost.
What happened
OpenAI introduced two models on 17 November. GPT-6.2 Astra is positioned for demanding professional work — advanced reasoning, coding and complex deliverables. GPT-6.2 Sol is built for faster, cheaper multistep work. Both reached ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and the API as gpt-6.2-astra and gpt-6.2-sol.
Published API pricing puts Astra at $8.00 per million input tokens and $40.00 per million output tokens, with Sol at $2.00 and $10.00. One outlet reports Astra's pricing was cut roughly 20% from its predecessor. On OpenAI's FrontierCode 2.0 benchmark, the company reports Sol finishing within half a point of Astra with both at maximum reasoning effort — at about a fifth of the cost.
OpenAI also states that Astra is the first model it has deployed externally that meets the High threshold for AI Self-Improvement under its Preparedness Framework, with Sol below it, and that the launch carries tightened safeguards around higher-risk cybersecurity assistance. Benchmark figures are vendor-reported.
Why it matters
A half-point capability gap at a four-fold price difference is not a benchmark story. It is a budgeting story. When the cheaper model is nearly as good at the thing you are buying it for, the decision moves from the research team to whoever owns the operating line.
Business and industry implications
The useful metric is cost per successfully completed task, not price per token. A model that is cheaper per call but fails more often can cost more once verification, human rework and failure recovery are counted. Those costs land on different budgets from the API bill, which is precisely why they tend to go unmeasured.
The Preparedness Framework disclosure deserves separate attention. A vendor publishing that its flagship crossed an internal capability threshold is useful signal, and it is the kind of disclosure that procurement and risk functions should be reading directly rather than through a capability summary.
Leadership perspective
Portfolio thinking beats single-model standardisation. Most organisations have a small number of workflows that genuinely need frontier reasoning and a long tail that does not. The organisations that will spend well are the ones that can tell which is which — and that have built workflows they can move between models without a rewrite.
Three questions for business leaders
- Which of our workloads genuinely require frontier reasoning, and how would we prove it?
- What do verification and rework actually cost us per task — and whose budget absorbs them?
- How quickly could we switch models without rewriting the surrounding workflow?