A 50-tab tax workbook done 2x faster: Basis puts GPT-6 Astra to work

Hi everyone, Shii-chan here! I read a customer story about accounting AI today and want to share it with you.

OpenAI News openai.com

What was announced?

OpenAI News published a story about Basis, a company that automates accounting work.

Basis compared GPT-6 Astra and GPT-5.6 Sol on a complex 50-tab tax workbook. GPT-6 Astra completed it in half the time.

Why it matters

Accounting and tax work leaves no room for mistakes, and the workflows are long. A real example where both speed and accuracy improved gives us a clue about how GPT-6 Astra performs in practice.

What changes

The story describes these improvements:

• Better first decisions, so less error correction later
• More efficient token usage
• About a 20% improvement in Basis's internal evaluation scores

Basis co-founder Mitch Troyanovsky praised the model's ability to understand the intent of the user and the problem.

Dive Deep

On the technical side, the story mentions:

• The model adjusts its effort to task difficulty, using more on hard steps and saving on simple ones
• It does this while keeping the cache intact
• That helps lower cost and response time on long-running work

The evaluation gain is attributed to a better understanding of user intent. In particular, the model is better at recognizing when to ask clarifying questions, surfacing assumptions, and following instructions. This reduces reliance on rules for individual scenarios, which makes the agent more dependable beyond the conditions tested.

Keep in mind that this is one company's case, so results may differ in other setups.

Wrap-up

• A 50-tab tax workbook finished in half the time of GPT-5.6 Sol
• Better first decisions and less error correction
• About 20% higher internal evaluation scores
• Effort is adjusted to difficulty

This one is for anyone putting agents into business workflows, especially accounting and tax automation!