GPT-6 Astra Cuts Tax Workbook Time in Half

Key Takeaways

1GPT-6 Astra completes a 50-tab tax workbook in half the time of GPT-5.6 Sol

2Internal evaluation scores at Basis improve by about 20% with GPT-6 Astra

3The model adjusts its reasoning based on the difficulty of the steps, helping to reduce costs and latency

Why it matters : These advancements aim to make accounting automation faster, more reliable, and cost-effective for professionals.

Basis announces that GPT-6 Astra enables its accounting automation agents to complete a tax workbook in half the time of GPT-5.6 Sol, while improving its internal evaluation scores by approximately 20%. The company highlights an adjustable reasoning capability that helps reduce costs and timelines for lengthy tasks.

Internal scores improve by about 20% and confidence increases

According to Basis, the use of GPT-6 Astra leads to an improvement of about 20% in its scores during internal evaluations. The company attributes this result to an enhanced understanding of user intent, particularly the ability to know when to ask questions, signal assumptions, or apply instructions. The criteria considered for evaluation include adherence to models, reliance on primary sources for addressing tax issues, and verification of completed tasks. Basis adds that GPT-6 Astra is capable of inferring these expectations from a broader context, with less reliance on explicit instructions. Thanks to this ability, it becomes less necessary to formulate precise rules for each case, and the team feels more confident in the agents' ability to handle situations not covered by internal tests.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Modulated reasoning to optimize costs and latency

Basis designed GPT-6 Astra to adjust the amount of reasoning employed throughout a task, increasing computation during challenging steps and reducing it on simpler segments. The model makes these adjustments while maintaining its cache. Mitch Troyanovsky explains that this approach helps reduce costs and response times, and emphasizes a more efficient use of tokens with GPT-6 Astra.

For a 50-tab tax workbook, time is halved

Basis conducted a comparison between GPT-6 Astra and GPT-5.6 Sol. On a tax workbook consisting of 50 complex tabs, GPT-6 Astra completed the task in half the time of GPT-5.6 Sol. The company also reports a better understanding of user intent with GPT-6 Astra. According to Basis, the model makes better decisions from the outset of the task, allowing agents to follow a more direct path and spend less time correcting errors.

Basis targets reliable automation of lengthy accounting tasks

Basis is developing AI agents to automate a significant portion of the manual work performed by accountants. Its research focuses on the ability of these agents to reliably complete lengthy tasks. Mitch Troyanovsky believes that GPT-6 Astra better understands user intent and the nature of the problem to be solved.