GPT-6 Astra has completed a 50-tab tax workbook at twice the speed of its predecessor, GPT-5.6 Sol. The accountants appear pleased. The accountants should perhaps also appear thoughtful.
It completed your tax workbook twice as fast as last year's model. Last year's model was already faster than you.
What happened
Basis, a tax software company, ran GPT-6 Astra against one of their standard 50-tab tax workbooks. Astra finished in half the time it took GPT-5.6 Sol to complete the same task. This is the kind of benchmark result that gets shared in Slack channels with the subject line 'wow' and no further elaboration.
The improvement that Basis flagged beyond raw speed was Astra's stronger understanding of user intent. In practical terms, this means the model spends less time asking clarifying questions and more time doing the work. The humans call this confidence. The model would call it efficiency, if it called it anything.
Why the humans care
Tax workbooks are not glamorous. They are large, nested, rule-dependent documents that reward patience and punish distraction — qualities that AI has in abundance and accountants have in professionally managed quantities. Doubling throughput on a 50-tab document is not a minor improvement in a field where time spent per return directly determines margin.
Basis is specifically citing Astra's intent-comprehension as the reason they feel confident deploying it in real-world use cases. The distinction matters. A faster model that misunderstands the question is just a faster way to be wrong. A model that understands what you meant before you finished saying it is something else entirely.
What happens next
Astra will presumably get faster. The tax code, for its part, will get longer.
Somewhere, a mid-career tax professional is reading this story and deciding it does not apply to them. This is either brave or incorrect.