Artificial intelligence news.
Straight from the original announcements.

Basis finishes 50-tab tax workbook in half the time with GPT-6 Astra

September 28, 2026By the AI News Bites Editorial TeamSource: OpenAI

Basis, a company that builds AI agents for accountants, says GPT-6 Astra completed a complicated 50-tab tax workbook in half the time of GPT-5.6 Sol. According to OpenAI's account of the test, the company also saw about a 20% improvement in its internal evaluation scores.

Key points
  • Basis says GPT-6 Astra completed a 50-tab tax workbook in half the time GPT-5.6 Sol needed.
  • Basis saw about a 20% improvement in internal evaluation scores.
  • GPT-6 Astra adjusts its reasoning effort per step while keeping its cache intact, which Basis says cuts cost and response time.
  • The model needs fewer explicit instructions, reducing the rules Basis must write for individual situations.

A 50-tab workbook in half the time

Basis builds AI agents that automate much of the manual work accountants do each day, with the aim of moving their time from repetitive tasks to strategic work. Its research focuses on agents that can reliably complete long tasks.

The company compared GPT-6 Astra and GPT-5.6 Sol on a complicated tax workbook with 50 tabs. The goal was to complete the workbook accurately and reliably, and GPT-6 Astra was markedly faster. Mitch Troyanovsky, co-founder of Basis, put it this way: "GPT-6 Astra is able to complete that workbook in half the time that GPT-5.6 Sol is able to."

Basis also noted that GPT-6 Astra makes better decisions at the start of a task. That helps its agents take a more direct path through the work, with less time spent correcting mistakes. Troyanovsky says this also makes the model more efficient in its use of tokens.

Reasoning that scales with the step

Basis has GPT-6 Astra adjust how much reasoning it uses as a task progresses. The model dials up computation when a step is difficult and uses less when a step is easier. According to the source, the model can make these adjustments while keeping its cache intact.

Troyanovsky says this approach reduces cost and response time. In his view, it makes long-running tasks more economical for Basis and its customers.

Better grasp of what accountants want

Basis reports about a 20% improvement in its internal evaluation scores with GPT-6 Astra. It attributes the gain to a better understanding of user intent, including when to ask questions, flag assumptions, and follow instructions. In Troyanovsky's words, "GPT-6 Astra does a better job of really understanding the intent of the user and the problem."

Basis evaluates both how its agents work and their final answers. The checks include whether agents follow templates, consult primary sources for tax questions, and check their own work.

The company says GPT-6 Astra can infer these expectations from a broader context, with fewer explicit instructions. That reduces the need for Basis to write rules for individual situations.

What it means for Basis and its customers

For Basis, the results point to faster completion of long accounting tasks and lower costs from more economical use of tokens and reasoning. The team also says it has more confidence that its agents can handle situations beyond those covered in internal tests.

The figures come from Basis's own testing and evaluations, as reported by OpenAI, and the source does not provide further detail on the methodology.

Related news

Research
ResearchNovember 15, 2026, 23:00

OpenAI outlines lattice construction for the full Standard Model

OpenAI has proposed a candidate nonperturbative definition of the Standard Model on a spacetime lattice, with GPT-6 Astra helping develop the construction and detailed…

Products and tools
Products and toolsNovember 13, 2026, 09:00

Shield hands legacy IT portal tasks to OpenAI-powered agents

Shield says its Forge platform can now hand supported browser tasks to AI agents powered by OpenAI’s GPT-6 Astra, allowing IT providers to automate work in legacy…

GPT-6 Astra led coding agents, but finished only 26.4% of tasks
ResearchNovember 10, 2026, 17:08

GPT-6 Astra led coding agents, but finished only 26.4% of tasks

OpenAI’s GPT-6 Astra passed an average of 69.9% of hidden acceptance criteria in a benchmark of coding agents, yet fully finished only 26.4% of tasks. The Co-Design…