Legora reviewed 41 documents in minutes with GPT-6 Astra
2026-09-04 · OpenAI
Legora Tests GPT-6 Astra on Financial Document Review
Legora has released results from a test that applied GPT-6 Astra to a financial document review workflow. The exercise was designed to evaluate the model across three dimensions: processing speed, error detection accuracy, and overall workflow performance relative to an established baseline.
Test Design
The workflow involved reviewing a set of 41 documents. To create a measurable accuracy benchmark, the team embedded four errors within the document collection before the review began. After the model completed its pass, results were checked against those planted issues, and overall performance was compared to the previous workflow standard.
Speed: 41 Documents in Minutes
GPT-6 Astra completed the full review of all 41 documents in a matter of minutes. Financial documents are typically dense, structured, and terminology-heavy — the kind of material that demands sustained attention from human reviewers and can take hours to work through at scale. Compressing that process to a few minutes represents a significant operational shift, particularly for teams that manage high-frequency or high-volume document workloads such as contract reviews, regulatory filings, or due diligence packages.
Accuracy: All Four Errors Found
The model identified every one of the four planted errors, achieving a perfect detection rate on that measure. In financial review settings, missing an error in a contract clause, a compliance disclosure, or a financial statement carries real consequences. Speed matters, but completeness is the harder requirement to meet consistently. The fact that GPT-6 Astra caught all four planted issues — with no misses — is the more demanding result to replicate, and it anchors the speed numbers with a meaningful accuracy guarantee.
Performance Improvement: Nearly 40%
Overall workflow performance improved by nearly 40% compared to the prior baseline. That aggregate figure reflects combined gains across the dimensions the test measured. A 40% improvement on a workflow that already handles professional-grade financial documents is a substantial margin, and it suggests that integrating GPT-6 Astra into this review process produces efficiency gains that go beyond incremental optimization.
What This Means
The Legora test adds a concrete, scenario-specific data point to the broader conversation about AI in professional document review. The results show that GPT-6 Astra, applied to a structured financial workflow, can compress review time while maintaining full accuracy on a planted-error detection task. For organizations weighing AI deployment in compliance, legal, or financial review functions, results at this level of specificity — a defined document set, a quantified accuracy measure, and a clear performance delta — provide a more actionable basis for evaluation than general capability claims alone.