Basis ran GPT-6 Astra against GPT-5.6 Sol on a 50-tab tax workbook. Astra finished the job twice as fast.
The speed gain is not the headline. What matters is intent comprehension. Astra understood what users actually meant, not just what they typed. That distinction is what gave Basis enough confidence to deploy it in a real-world accounting workflow, where misreading a prompt has financial consequences.
The full case study breaks down how the benchmark was structured across all 50 tabs and where Sol specifically failed on ambiguous inputs. If you work in finance, legal, or any domain where precision is non-negotiable, the methodology section alone is worth your time.
[READ ORIGINAL →]