Enterprise work rarely comes down to a single question. For an AI agent to handle enterprise work, it needs to be able to reason across many types of content and produce a finished result that holds up to review: a recalculated gradebook, a corrected drug regimen, an investment memo with every required sign-off. That’s what the Box AI Complex Work Eval measures. Each task hands an agent a realistic bundle of enterprise content, including spreadsheets, PDFs, presentations and images, and asks for the finished work product.We grade outputs against detailed, weighted rubrics and run tasks multiple times.We evaluated Google's Gemini 4 Argon against Gemini 3.1 Pro, and saw a substantial step up. Gemini 4 Argon scored 85% against 71% for Gemini 3.1 Pro, a 14 percentage point gain.
Gains across industries
The gains are largest in the most technical, evidence-heavy industries, where the answer depends on reading the source material closely and following it all the way through.
Public sector (76% to 94%): One task asked an agent to recalculate every student's final grade under a new weighting policy, and one assignment column carried an ambiguous label. Gemini 4 Argon worked out the column's correct category from the score patterns in the data, so every recalculated grade and the top-student ranking came out right. Gemini 3.1 Pro tended to put the column in the wrong category, and the error flowed through every grade it recalculated.
Life sciences (61% to 88%): In a drug-target identification task, the deliverable needed figures from a protein-interaction network that appears only as a diagram in the source report. Gemini 4 Argon read the network directly from the figure. Gemini 3.1 Pro answered that step from an adjacent table describing a different subset, which gave it the wrong network counts.
Healthcare (57% to 87%): A radiology quality-control review asked the agent to audit a full batch of imaging studies against their written reports. Gemini 4 Argon worked through every study in the batch. Gemini 3.1 Pro typically reviewed only part of it, and in one run reported a finding that wasn't in the source material.
Legal (64% to 75%): Drafting an acquisition investment memo meant following the firm's internal guidelines on who has to sign off before the memo goes to final review. Gemini 4 Argon read the guidelines all the way through and caught that a later section adds the product and compliance leads to the required approvers. Gemini 3.1 Pro stopped at the first approval rule and under-counted the sign-offs.
Media and entertainment (65% to 79%): A casting review for a bilingual ad campaign asked the agent to evaluate every voice actor in the pool against five requirements: union status, fluency in both languages, accent, the role's fee range and fit for the brand. Gemini 4 Argon worked through every requirement for every candidate, in both roles, and handed back a complete eligibility record. Gemini 3.1 Pro stopped at each candidate's first disqualifier, so most of the evaluation the brief asked for never got done.
Results by type of work
By mode of work, Gemini 4 Argon's biggest gain is 23 points in data analysis, followed by 15 points in drafting reports from source data. Both modes test whether a model can carry a multi-step analysis from raw source material all the way to a finished, internally consistent deliverable.
Take a management plan for an unstable intensive-care patient. Both models spotted the dangerous potassium level and a serotonin-toxicity interaction between drugs. Only Gemini 4 Argon went back to the medication list and removed the drugs causing both, so its final regimen matched its own diagnosis. The pattern repeats in a product-launch evaluation, where Gemini 4 Argon saw that existing inventory could bridge an early capacity gap. This led to it recommending an earlier launch and a faster break-even.
What this means for enterprise work
Gemini 4 Argon carries the work through to the end. It reads the figure as well as the table, reviews every item it's asked to check instead of stopping partway, follows the guidelines to the last section, and acts on its own diagnosis. Those are exactly the habits that separate a deliverable you can hand off from one you have to check line by line. It also has a tighter worst case on run time, which matters when an agent is running recurring work in the flow of someone's day. For teams putting models on complex, document-heavy analysis, Gemini 4 Argon is a clear step up from Gemini 3.1 Pro.
Gemini 4 Argon is coming soon to Box AI.
