Basis, a company that builds AI agents to automate routine accounting and tax tasks, reports notable gains in speed and internal evaluation scores after adopting GPT‑6 Astra. The firm tested the models on a complex tax workbook composed of 50 tabs, where the objective was to complete the workbook accurately and reliably.
Time savings and performance
According to Basis, GPT‑6 Astra finished the same task in about 50% less time than GPT‑5.6 Sol. The company also recorded approximately a 20% improvement in its internal evaluation scores when using Astra.
Better inference of user intent and decision‑making
Basis says GPT‑6 Astra demonstrates a stronger understanding of what users intend to accomplish: it asks clarifying questions at appropriate times, flags assumptions, and follows instructions more reliably. Those improved initial decisions let the agents follow a more direct path through the work with fewer corrections needed.
Dynamic allocation of reasoning and cost efficiency
Basis further reports that Astra can adjust how much reasoning effort it applies as a task progresses—ramping up computation for difficult steps and using less for simpler ones—while retaining its cache. This behavior reduces token consumption, shortens response times, and makes long‑running tasks more economical for Basis and its customers.
Fewer handcrafted rules and greater confidence in generalization
The company evaluates both agent workflows and final outputs, checking whether agents adhere to templates, consult primary sources on tax questions, and verify their own work. GPT‑6 Astra can infer these expectations from a broader context with fewer explicit instructions, lowering the need for Basis to write rules for individual situations and increasing confidence that agents will handle cases beyond internal tests.
Summary
Basis's internal testing shows that GPT‑6 Astra delivers roughly a 50% reduction in completion time for a 50‑tab tax workbook and about a 20% uplift in evaluation scores compared with GPT‑5.6 Sol. The model's improved intent understanding, dynamic reasoning allocation, and better initial decisions reduce manual oversight and operational cost for long tasks.
