Basis cuts tax workbook completion time by 50% with GPT-6 Astra

Basis, an AI agent startup for accountants, reports that GPT-6 Astra completes a complex 50-tab tax workbook in 50% less time than the previous model, GPT-5.6 Sol. The company attributes this speed to the new model's ability to make better initial decisions, allowing agents to take a more direct path through tasks with fewer corrections.

Improved Efficiency and Intent Understanding

Mitch Troyanovsky, Co-founder of Basis, notes that GPT-6 Astra demonstrates a deeper understanding of user intent. This capability allows the model to infer expectations from broader context, reducing the need for explicit instructions and helping agents know when to ask questions or flag assumptions.

The model also supports dynamic reasoning effort, adjusting computation levels based on task difficulty while keeping its cache intact. This feature helps reduce costs and response times for long-running tasks.

Evaluation Results

In Basis’s internal evaluations, GPT-6 Astra showed an approximate 20% improvement in scores. These improvements were driven by better adherence to templates, consultation of primary sources, and self-checking behaviors, giving the team more confidence in handling scenarios beyond their internal test sets.