Basis Tax Automation with GPT-6 Astra

Basis has integrated GPT-6 Astra to automate manual accounting tasks, resulting in a 2x increase in speed for completing complex tax workbooks and a 20% improvement in internal evaluation scores. This performance gain is driven by the model's improved understanding of user intent and more efficient task execution.

Performance Gains in Complex Tax Workbooks

GPT-6 Astra completes a complicated tax workbook consisting of 50 tabs in half the time required by GPT-5.6 Sol. According to Mitch Troyanovsky, Co-founder of Basis, the model demonstrates a stronger understanding of user intent and the specific problems accountants aim to solve.

This speed increase is attributed to better decision-making at the start of a task. By identifying the correct path more quickly, the agents spend less time correcting mistakes and use tokens more efficiently.

Dynamic Reasoning and Resource Optimization

Basis utilizes GPT-6 Astra's ability to adjust reasoning effort based on the difficulty of the task. The model can increase computation for difficult steps and decrease it for easier ones while maintaining its cache. This dynamic adjustment reduces both response times and operational costs, making long-running accounting tasks more economical.

Improvements in Reliability and Intent Understanding

Basis observed a 20% improvement in its internal evaluation scores. This improvement is driven by several key capabilities:

• Intent Recognition: Better understanding of when to ask clarifying questions, flag assumptions, and follow specific instructions.

• Contextual Inference: The model can infer expectations—such as following templates, consulting primary tax sources, and self-checking work—from broader context with fewer explicit instructions.

• Reduced Rule Dependency: Because the model can infer expectations, Basis has reduced the need to write explicit rules for individual scenarios, increasing confidence in the agent's ability to handle edge cases not covered in internal tests.