Basis Completes Tax Workbook 2X Faster With GPT-6 Astra

Accounting automation firm Basis reports that its AI agents completed a complex 50-tab tax workbook in half the time using OpenAI’s GPT-6 Astra compared to the previous GPT-5.6 Sol model. The company, which builds agents to handle manual accounting tasks, noted significant gains in both speed and reasoning efficiency.

What Happened

Basis conducted a comparative analysis between GPT-6 Astra and GPT-5.6 Sol on a complicated tax workbook containing 50 tabs. The primary objective was to complete the workbook accurately and reliably. Mitch Troyanovsky, co-founder of Basis, stated that GPT-6 Astra completed the task in half the time required by its predecessor. The company also observed that the new model makes better decisions at the onset of tasks, allowing agents to take more direct paths and spend less time correcting errors. This efficiency extends to token usage, which Troyanovsky says is lower with the new model.

Why It Matters

The performance improvements suggest that GPT-6 Astra offers enhanced contextual understanding for long-running professional tasks. Basis reported approximately a 20% improvement in its internal evaluation scores, driven by the model’s ability to better understand user intent, including when to ask clarifying questions or flag assumptions. The model also demonstrates dynamic reasoning adjustment, dialing up computation for difficult steps and reducing it for easier ones while maintaining cache integrity. This capability reduces operational costs and response times for long tasks, potentially making automated accounting more economical for Basis and its customers. Additionally, the model’s ability to infer expectations from broader context reduces the need for explicit, rule-based programming for individual scenarios.

The Bottom Line

Basis’s internal tests indicate that GPT-6 Astra significantly accelerates complex accounting workflows and improves agent reliability through better intent understanding and adaptive reasoning, though these claims remain specific to Basis’s internal evaluation metrics.