AI-generated editorial still life of financial documents, translucent comparison sheets, a calculator and a pencil.
AI-generated editorial illustration. No actual test results are shown.

The part I want to understand

Looking at a finance benchmark leaves me with a practical question: what would make an AI answer useful enough to enter my model? A score helps me decide what to investigate. I still need to know whether the answer uses the right reporting period, the right units and a source I can check.

As a finance student interested in processing financial data, I would like to start with a small Gemini workflow and then repeat the same task with another accessible model. The part that interests me is the work between receiving an answer and being comfortable using it. That is also where an apparently quick response could leave a lot of manual checking.

One task I could actually check

In the October 12 Vals update, Gemini 4 Argon scores 65.40%, Wonder Full 59.06% and Wonder 56.64%. These are weighted partial-credit scores, not percentages of complete financial analyses performed correctly. The test has 450 tasks, three runs per model and three model judges. A failed critical check gives a task zero credit; the separate All-Pass measure requires every check to pass. I have not verified current All-Pass results for these models.

For my own exercise, I would choose one company and two annual filings. I would ask each model to extract revenue and operating income, identify the period and units, and point to the exact table or note. Then I would calculate operating income divided by revenue in Excel, keeping the reported figures separate from formulas. If the earlier year had been restated, I would record which version I used and why.

I would also ask what makes the comparison misleading: a changed segment boundary, a disposal or a non-GAAP adjustment could matter more than a small arithmetic difference. Agreement between two models would be a reason to inspect the evidence, not independent confirmation.

There are practical limits. Meta says Wonder is available through its API and Muse Code, while Full is offered to Muse Code subscribers. I would check my actual access and record the tools and settings before comparing anything. A consumer app exercise would not reproduce Vals’ shared testing environment, and the benchmark alone does not establish a statistically significant advantage for my task.

What I would try next

I would first complete the Gemini exercise with a source log and an Excel check, then freeze those inputs and repeat it with the alternative. My comparison would record missing citations, wrong units or periods, corrections, review time and measured usage cost. I have not run this experiment yet.

This is the skill I want to build: making a financial answer traceable enough that I can explain why I accepted it. Once that works for one company, I would explore using n8n to collect new filings into a review queue. The next question is how a revised filing should change an approved input or assumption without silently overwriting the earlier reasoning.

Sources