Beyond Extraction: How Five Models Handle Cross-Document Synthesis
GPT-6.1 Terra, GPT-6.1 Sol, GPT-5.6 Terra, Sonnet 5.5, and Opus 5.5
In Private Equity due diligence, many questions span multiple filings. A deal team may need to track a margin across five years of annual reports, or gather every disclosure on a topic. Answering them takes cross-document synthesis: finding the right evidence across documents and combining it into one answer.
These questions rarely have one clean number to grade. A useful answer also has to cover the full question, stay consistent across repeated runs, and point back to its sources.
What we tested
The ToltIQ Research Team ran GPT-6.1 Terra, GPT-6.1 Sol, GPT-5.6 Terra, Sonnet 5.5, and Opus 5.5 through the same 10 prompts, three runs each, inside a data room of more than 1,000 public filings. We compared the responses on consistency, completeness, document grounding, and response time.
Note: The ToltIQ Research Team tests frontier models as they are released, whether or not we offer them on our platform. For the models we offer and why, read Evaluating AI Models for Private Equity Due Diligence.
What stood out
Consistency, completeness, and speed rarely come together in one model. The fastest of the five was also the least consistent. And when a model’s figures changed between runs, it was usually because the model changed a definition, such as switching between fiscal and calendar quarters, not because it miscalculated.
Download the full report to see how each of the five models performed, the median response time for each, and a side-by-side example of two models answering the same prompt.
Share this article