
OpenAI has introduced GPT-6.2 Astra and GPT-6.2 Sol, two models aimed at different kinds of work. Astra is positioned for demanding professional projects; Sol is the more efficient option for everyday tasks. The announcement is as much about the cost of completing work as it is about benchmark scores.
OpenAI announced the models on 17 November 2026. It says they are available through ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through its developer API. Eligible Enterprise and Edu access still depends on workspace administrators enabling it. Availability should not be read as unlimited use on every account.
What separates Astra and Sol?
OpenAI describes Astra as its frontier model for difficult projects, including documents, spreadsheets, presentations, code and interactive design. It says Sol improves on GPT-6.1 Sol for routine technical and non-technical work, with an emphasis on efficiency and collaboration.
These are the supplier’s descriptions, not results from testing conducted for this article. Examples selected for a launch can demonstrate a capability without establishing that every document, codebase or design brief will receive the same quality of output.
Lower token prices are only part of task cost
Standard API prices for Astra are US$8 per million input tokens, US$0.40 per million cached input tokens and US$40 per million output tokens, according to the announcement. OpenAI says its input and output token prices are 20% below GPT-6 Astra, with a deeper reduction for cached input.
Sol retains GPT-6.1 Sol’s standard API rates: US$2 per million input tokens, US$0.10 for cached input and US$10 for output. These are API rates, not a new price list for consumer subscriptions.
A lower token price does not guarantee the same reduction on every finished job. Total cost also depends on how many tokens a task uses, what can be cached, whether work needs retries and the services surrounding the model. The company’s claims about lower cost per task therefore need to be read alongside the evaluation conditions, not substituted for a budget for a real workflow.
What benchmark gains do, and do not, establish
OpenAI reports improvements on professional-work, coding, science and design evaluations. Those results can help identify tasks worth testing, but they do not establish universal superiority. A benchmark measures selected tasks under particular conditions; a business may have different documents, error tolerances and review requirements.
A practical comparison would use the same representative brief, source material and acceptance checks for both models. Review the answer’s factual accuracy, whether files actually work and how much correction is needed, as well as speed and cost. A polished presentation is not evidence that its underlying numbers are right.
Human checks still matter
The launch does not remove the need to verify citations, calculations, code changes or consequential recommendations. Nor does better performance on a test mean that a model can safely be given unrestricted access to private data or permission to take irreversible actions.
The useful takeaway is a new pair of options to evaluate: Astra for harder projects and Sol for efficient everyday work, as positioned by OpenAI. The choice for a particular organisation still needs evidence from its own tasks. All performance claims in this article are attributed to the company; no independent benchmark testing was carried out here.
Image: original concept illustration, not a product screenshot or a depiction of physical model hardware. Prices and availability checked against the 17 November announcement.