Accounting agent developer Basis reported that GPT-6 Astra completed complex tasks 50% faster than GPT-5.6 Sol. In a specific test involving a 50-tab tax workbook, the model demonstrated markedly higher speed and a deeper understanding of user intent.
Model ReleasesBasisGPT-6 Astra
Basis reports significant speed improvements using GPT-6 Astra for complex tax workbooks
According to Basis, the model's ability to make better decisions at the start of a task allows agents to take more direct paths, reducing the time spent on error correction and making token usage more efficient. The model also adjusts its reasoning computation dynamically based on task difficulty—increasing computation for difficult steps and decreasing it for easier ones—while maintaining its cache. This capability helps reduce both cost and response time for long-running tasks.
Basis's internal evaluation scores improved by approximately 20% with the new model. This improvement was driven by the model's ability to follow instructions, flag assumptions, and ask questions by inferring expectations from broader context. This capability reduces the need for the development team to write specific rules for every individual situation.
Sources
- Basis completes a tax workbook 2x faster with GPT-6 Astra （OpenAI News, 2026-09-28）
- Basis