Basis completes a tax workbook twice as fast with GPT-6 Astra

OpenAI has published a customer story describing how Basis, a tax automation company, used its new GPT-6 Astra model on a real tax workbook. According to the account, the workbook ran to 50 tabs, and GPT-6 Astra completed it twice as fast as GPT-5.6 Sol.

Speed was not the only point the case makes. OpenAI says the newer model showed a stronger grasp of user intent, which gave Basis more confidence that it would hold up in day-to-day work rather than only in a controlled demo.

A 50-tab workbook is a meaningful stress test because it is not a single question and answer. It asks the model to keep cross-sheet references, domain conventions and internal consistency intact, when a single wrong cell can make the whole output unusable.

The shape of the comparison also matters. The case puts Astra and Sol on the same document instead of pointing to an abstract leaderboard. For enterprise buyers, that kind of same-task comparison is often more informative, because it resembles the file and the workflow they actually have.

One caveat deserves to stay in view: this is still a customer story published by the model vendor, with figures supplied by OpenAI and its customer and not yet reproduced by a third party, so procurement teams should read it alongside public benchmarks and their own pilot results.

What to watch next is whether more customers publish comparable same-task comparisons, and whether accuracy-critical fields such as tax, audit and finance become the first places where the new generation of models is validated in everyday work.