Legora ran GPT-6 Astra through a financial document review workflow covering 41 documents. The model found all four deliberately planted errors. Performance improved by 39%.
The number that matters is not the speed, it is the error recall. Four hidden errors, zero missed. That kind of precision in financial document review has direct liability implications, and the workflow mechanics behind how Legora structured the prompts and validation steps are what make this worth reading in full.
This is an early benchmark from a production legal-tech environment, not a lab test. As GPT-6 Astra access expands, expect more firms to publish similar runs. The gap between these results and what human reviewers produce at the same volume is the real story OpenAI is building toward.
[READ ORIGINAL →]