Basis AI Speeds Tax Work 2x with GPT-6 Astra

Overview

Basis, a startup specializing in AI agents for accounting automation, has reported a significant performance leap with OpenAI’s latest model. By integrating GPT‑6 Astra, the company claims its agents can now complete complex tax workbooks approximately twice as fast as they could with the previous generation, GPT‑5.6 Sol.

This speed increase is not merely a matter of raw processing power but is attributed to a deeper, more nuanced understanding of user intent. For an industry where accuracy and compliance are paramount, the ability to interpret accountant objectives with greater fidelity represents a critical advancement in automated workflow tools.

"GPT-6 Astra does a better job of really understanding the intent of the user and the problem." — Mitch Troyanovsky, Co-founder of Basis

The announcement underscores a broader trend in the enterprise AI sector: the shift from simple task execution to strategic collaboration. As Basis continues to refine its research into reliable, long-running agents, the integration of Astra signals a maturing capability to handle the intricate, multi-step processes that define professional accounting.

Performance Benchmarks and Efficiency Gains

The core of Basis’s recent testing involved a demanding real-world scenario: a comprehensive 50-tab tax workbook. These workbooks are notoriously complex, requiring agents to navigate intricate financial data, adhere to strict formatting templates, and ensure consistency across numerous interconnected sheets. Previously, using GPT‑5.6 Sol, agents would spend considerable time interpreting requirements and correcting errors mid-process.

With GPT‑6 Astra, Basis recorded a 50% reduction in time required to finish this same workbook. This dramatic improvement suggests that the new model reduces the overhead associated with navigation and correction. In practical terms, what once took hours of computational resources and human oversight can now be accomplished in significantly less time, allowing accountants to pivot from manual oversight to high-value strategic analysis much sooner.

Beyond raw speed, the efficiency gains extend to token usage. Because the model makes better decisions at the outset of a task, it tends to take a more direct path toward completion. This reduces the number of erroneous turns or clarifying loops that often plague earlier model versions. Consequently, Basis reports that the model is not only faster but also more economical in its operation, lowering the cost per task for both the company and its end users.

Adaptive Reasoning and Cost Management

A distinctive feature of GPT‑6 Astra highlighted by Basis is its ability to dynamically adjust its reasoning effort. Unlike static models that apply uniform computational intensity to every step, Astra can assess the difficulty of a specific task component in real time. When encountering a complex calculation or a ambiguous regulatory detail, the model increases its computational focus. Conversely, for straightforward data entry or verification, it conserves resources.

Crucially, the model can make these adjustments while maintaining its context cache. This capability ensures that long-running tasks remain coherent throughout their duration, preventing the "lost in the middle" phenomenon where earlier instructions fade from relevance. For Basis’s agents, which may operate continuously over long periods, this stability is essential for maintaining accuracy.

This adaptive approach has a direct impact on cost and response time. By dialing down computation when the task allows, the model reduces latency and operational expenses. Troyanovsky noted that this efficiency makes long-running autonomous tasks far more viable for commercial deployment. It transforms AI agents from novelty prototypes into reliable workhorses capable of handling sustained, complex workloads without prohibitive costs.

Enhanced Accuracy and Internal Evaluations

While speed is a compelling metric, reliability remains the primary concern for any tool deployed in financial services. Basis conducted extensive internal evaluations to measure the qualitative improvements brought by GPT‑6 Astra. The results showed an approximate 20% improvement in internal evaluation scores compared to previous benchmarks.

This improvement is driven largely by the model’s enhanced ability to infer expectations from broader contexts. In previous iterations, Basis had to write explicit, granular rules for individual situations to ensure agents followed specific templates or consulted primary tax sources. With Astra, the agents can deduce these requirements from general instructions, reducing the need for hard-coded logic.

The evaluation criteria included whether the agents:

• Strictly adhered to provided formatting templates.
• Consulted primary sources for ambiguous tax questions.
• Performed self-checks to verify their own outputs before finalizing.

GPT‑6 Astra demonstrated a superior capacity to follow these protocols without exhaustive pre-programming. This suggests a higher degree of robustness, giving Basis confidence that its agents can generalize well beyond the specific scenarios covered in internal tests.

Implications for the Accounting Workflow

For Basis, located in the technology sector within North America, these advancements have direct implications for their customer base. The company’s mission is to automate the manual drudgery of accounting, thereby freeing professionals to focus on strategic decision-making. The proven ability of GPT‑6 Astra to halve workbook completion times accelerates this transition.

Accountants can now expect faster turnaround times for routine filings and complex audits. The reduced need for manual correction and the model’s proactive approach to flagging assumptions and asking clarifying questions further enhance the trustworthiness of the output. As Basis continues to build on this foundation, the integration of more capable models like Astra will likely expand the scope of tasks that can be safely automated, pushing the boundaries of what AI agents can achieve in regulated industries.