Editorial illustration for Basis Finishes Complex Tax Workbook Twice as Fast Using GPT-6 Astra
GPT-6 Astra Cuts Tax Workbook Time in Half
Basis Finishes Complex Tax Workbook Twice as Fast Using GPT-6 Astra
Basis, a startup building AI agents for accounting work, ran a 50-tab tax workbook through two versions of OpenAI's models to see how much faster the newer one could get the job done. The result: GPT-6 Astra finished the workbook in half the time it took GPT-5.6 Sol, according to the company. Basis also reported roughly a 20% jump in its internal evaluation scores when testing Astra against Sol.
The company's agents are meant to take over the repetitive parts of an accountant's day, tasks like reconciling entries across dozens of tabs, so accountants can spend more time on judgment calls and client strategy. Basis's research has centered on getting agents to handle long, multi-step tasks without losing accuracy along the way, and the workbook test was designed to stress that exact capability.
Part of the speed gain, Basis found, comes down to how Astra plans its approach before it starts working. Better early decisions mean fewer wrong turns later, which cuts down on the time agents spend backtracking and fixing mistakes. Kirill Troyanovsky, who worked on the comparison at Basis, described what changed in the model's behavior.
Basis compared GPT‑6 Astra and GPT‑5.6 Sol on a complicated tax workbook with 50 tabs. The task was to complete the workbook accurately and reliably, and GPT‑6 Astra was markedly faster.
Why this matters
A 50-tab tax workbook finished in half the time isn't a benchmark curiosity, it's the kind of task that decides whether accounting firms actually trust an agent with client work during crunch season. Basis picked a brutal test case on purpose: tax prep punishes any model that guesses at intent instead of tracking it across dozens of interdependent tabs. That the gains showed up both in the workbook test and in Basis's broader internal evals (roughly 20%) suggests something more than a one-off tuning win, though we'd want to see how Astra handles a messier, less-curated dataset before calling it settled.
For founders building vertical agents, the lesson is less about GPT-6 Astra itself and more about what Basis is optimizing for: intent-tracking on long, structurally rigid documents. That's a narrower, more testable claim than "smarter model," and it's the kind of claim other AI-for-professional-services startups will now have to match or explain away. Watch whether Basis publishes error rates alongside speed, because a faster wrong answer in a tax filing is worse than no answer at all.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Anthropic's Claude Gains Ground With Focus on Operational Knowledge
From Archives Researchers Find 'Skill Cascading' Attack Threatens AI Agents
Common Questions Answered
How much faster did GPT-6 Astra complete the tax workbook compared to GPT-5.6 Sol?
GPT-6 Astra finished the complex 50-tab tax workbook in half the time it took GPT-5.6 Sol to complete the same task. This significant speed improvement demonstrates a substantial performance gain for accounting automation workflows.
What was the performance improvement shown in Basis's internal evaluation scores when testing GPT-6 Astra?
Basis reported roughly a 20% jump in its internal evaluation scores when testing GPT-6 Astra against GPT-5.6 Sol. This improvement across multiple evaluation metrics suggests that Astra's advantages extend beyond just speed to overall accuracy and reliability.
Why did Basis choose a 50-tab tax workbook as their test case for comparing these AI models?
Basis deliberately selected a complex 50-tab tax workbook because tax preparation punishes any model that guesses at intent instead of accurately tracking information across dozens of interdependent tabs. This challenging test case reflects real-world accounting work and determines whether firms can actually trust an AI agent with client work during busy tax seasons.
What types of repetitive accounting tasks is Basis's AI agent designed to handle?
Basis has built AI agents specifically designed to take over the repetitive parts of an accountant's day, automating routine tasks that consume significant time during tax preparation and other accounting work. The company's focus on automation aims to free accountants from mundane tasks so they can focus on more complex client work.
Further Reading
- Basis completes a tax workbook 2x faster with GPT-6 Astra - OpenAI
- Basis scales accounting by turning OpenAI model progress into autonomous workflows - OpenAI
- Basis | Oleksandr Zalizniak - Oleksandr Zalizniak
- Basis | Oleksandr Zalizniak - Oleksandr Zalizniak
- Basis completes tax workbook 2x faster with GPT-6 Astra - AI Pricing Guru
