Basis completes a tax workbook 2x faster with GPT-6 Astra
Accounting startup Basis says GPT-6 Astra finishes a 50-tab tax workbook twice as fast as GPT-5.6 Sol.
Basis, which builds AI agents that automate accounting work, said GPT-6 Astra completed a 50-tab tax workbook in half the time of GPT-5.6 Sol. Co-founder Mitch Troyanovsky said Astra better infers user intent, chooses a more direct path at the start, and uses tokens more efficiently. Basis also has the model raise or lower reasoning effort during a task while preserving its cache, which it says reduces cost and latency. Internal evaluation scores improved about 20%, covering template use, consultation of primary tax sources, and self-checking.
- GPT-6 Astra finished a 50-tab tax workbook in half GPT-5.6 Sol's time.
- Basis reported about a 20% gain on internal agent evaluations.
- Astra varies reasoning effort mid-task while keeping its cache intact.
- Better intent understanding cut extra rules and early-path mistakes.
Full article362 words · extracted from openai.com · click to collapse
Basis(opens in a new window) builds AI agents to automate much of the manual work that accountants do each day, helping them shift their time from repetitive tasks to strategic work. The company’s research focuses on agents that can reliably complete long tasks, and with GPT‑6 Astra, it’s seeing a stronger understanding of what accountants want to accomplish.
“GPT-6 Astra does a better job of really understanding the intent of the user and the problem.”
—Mitch Troyanovsky, Co-founder, Basis
Completing a 50-tab tax workbook in half the time
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.
“GPT-6 Astra is able to complete that workbook in half the time that GPT-5.6 Sol is able to.”
—Mitch Troyanovsky, Co-founder, Basis
Basis also noted that GPT‑6 Astra makes better decisions at the start of a task, helping Basis’s agents take a more direct path through the work with less time spent correcting mistakes. Troyanovsky says that also makes the model more efficient in its use of tokens.
Matching reasoning effort to the task
Basis also has GPT‑6 Astra adjust how much reasoning it uses as a task progresses, dialing up computation when a step is difficult, and using less when a step is easier. The model can make these adjustments while keeping its cache intact. Troyanovsky says this helps reduce cost and response time, making long-running tasks more economical for Basis and its customers.
Building confidence in real-world use
Basis saw about a 20% improvement in its internal evaluation scores with GPT‑6 Astra, driven by better understanding of user intent, including when to ask questions, flag assumptions, and follow instructions.
Basis evaluates how its agents work and their final answers, including whether they follow templates, consult primary sources for tax questions, and check their own work. 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 and gives the team more confidence that its agents can handle situations beyond those covered in internal tests.
Text extracted automatically; images, tables and formatting may be missing. Original: