OpenAI on November 17, 2026, announced GPT-6.2 Astra, a new model designed for complex professional work, and GPT-6.2 Sol, an updated model aimed at everyday tasks. The company said both models are intended to deliver stronger results while using fewer output tokens, a factor that can affect the cost of operating AI systems at scale.
OpenAI said GPT-6.2 Astra outperformed its predecessor on several evaluations while reducing the cost of completing the tested work. On Agents’ Last Exam, a benchmark for multistep professional tasks, Astra scored 5.2 points higher than GPT-6 Astra and cost less than one-third as much, according to the company. OpenAI also said Astra finished 5.5 points ahead of Anthropic PBC’s Claude Fable 5.2 on that evaluation at 5% of the cost.
A model aimed at professional workflows
The company is positioning Astra for work in areas including science, law, finance, accounting and sales. OpenAI said the model scored 9 percentage points higher than GPT-6 Astra on Terminal-Bench-Science 0.1, an evaluation of scientific work performed in a terminal. On Harvey LAB-AA, which tests legal assignments, the company said Astra more than doubled the pass rate of Fable 5.2 at approximately one-sixth of the cost.
Beyond benchmark results, OpenAI said Astra was trained to produce slides, spreadsheets and documents that better match professional conventions. The model can reproduce a slide template from an original file or screenshot, including its layout, colors and typography, down to the exact font size, according to the company.
- GPT-6.2 Astra: aimed at demanding professional work, including scientific and legal assignments.
- GPT-6.2 Sol: positioned as OpenAI’s model for everyday work and higher task volume on the same paid plan.
- Professional output: OpenAI said Astra improves the formatting and structure of slides, spreadsheets and documents.
Early business feedback and the everyday model
Accounting AI startup Basis, an early tester, said its average time to complete long-horizon accounting work fell by 50% with Astra. The startup also reported a reduction of up to 73% for its most complex work. Those results come from an early tester and do not establish performance across all accounting tasks or organizations.
OpenAI described GPT-6.2 Sol as its general model for everyday work and said users could complete more tasks on the same paid plan than with GPT-6.1 Sol. The company also said Sol came within 0.5 points of Astra on Cognition AI Inc.’s FrontierCode 2.0 coding benchmark while costing about one-fifth as much.
The release reflects the growing focus among AI providers on the cost of completing extended tasks, rather than only the price or performance of a single response. OpenAI’s announcement combines claims about benchmark gains, document production and business workflow efficiency, while the reported figures remain tied to the company’s evaluations and the experience of named early testers.