Basis completes tax workbook twice as fast with GPT‑6 Astra

The company says GPT‑6 Astra better understands user intent, asks clarifying questions, flags assumptions and follows instructions with fewer explicit rules. Troyanovsky notes the model can adjust its reasoning effort during a task, increasing computation for difficult steps and scaling back for easier ones while keeping its cache intact, which lowers token usage, cost and response time.

Basis believes these advances let its agents take more direct paths through long‑running tasks, reducing the need for manual rule‑writing and improving confidence that agents can handle situations beyond internal tests.