Box CEO Aaron Levie says Gemini 4 Argon scored 14 percentage points higher than Gemini 3.1 Pro in Box AI Complex Work's internal evaluation. The attached chart reports full-dataset accuracy of 71% versus 85% and results for five industry subsets. This is a Box-run evaluation; the public material does not disclose the sample size or complete scoring and call configurations, so it should not be treated as an independent or readily reproducible general-purpose benchmark.
Useful for: Understanding Box's first-party comparison in its evaluation of complex enterprise-content tasks, along with the document-analysis examples shown in the post.
Not suitable for extrapolating to: Accuracy on agents, document collections, prompts, or industry tasks outside Box; or making medical, financial, or other professional decisions.
Model versions covered: Gemini 4 Argon and Gemini 3.1 Pro; specific API IDs, snapshots, and reasoning levels are not given.
Test environment or client: Box AI Complex Work eval; Box Agent (the post explicitly says Argon was evaluated through Box Agent). The chart footnote identifies the dataset as the Box AI V18 evaluation set, covering hard reasoning tasks on enterprise content.
Reasoning level and parameters: Not specified.
Levie says Box tested Gemini 4 Argon with its Box AI Complex Work eval one week before the release. The tasks involved complex work across enterprise documents from multiple industries, intended to approximate time-consuming tasks knowledge workers handle with enterprise content in their day-to-day work. The post compares Argon with Gemini 3.1 Pro and says the overall result was “up 14 percentage points across all tests.”
The attached image shows two sets of accuracy results: the Full Dataset and Industry Subset. Its footnote describes hard reasoning tasks on enterprise content in the Box AI V18 evaluation set. The post also lists five illustrative winning tasks, with reported margins of +18, +25, +37, +16, and +33 points. It does not explain how those example scores relate to the chart's accuracy figures, so the two sets of results are recorded separately.
The five illustrative task results listed in the post:
| Industry and margin stated by the author | Task and reported difference in the post |
|---|---|
| Financial services (+18 points) | A family business cash-flow analysis required laying out projected statements, loan interest, and annual overdrafts line by line. The author says Argon built and calculated the cash shortfall line by line as requested, while Gemini 3.1 Pro combined the interest into a single figure. |
| Healthcare (+25 points) | In an intensive care management plan, both models identified dangerous potassium levels and drug interactions. The author says only Argon then removed the relevant medication from the final plan, making it consistent with the diagnosis. This records the evaluator's description of consistency in the output; it is not clinical advice. |
| Technology (+37 points) | A product performance review asked which of two regional markets should be prioritized for cloud infrastructure optimization. The author says Argon selected the correct data for each region's monthly users, retention rate, and data per user, then chose North America. Gemini 3.1 Pro often pulled the wrong rows, at times reported retention above 100%, and pointed to the wrong region. |
| Energy (+16 points) | Energy reports for two facilities required identifying gaps in meter data. The author says Argon consistently noticed that smart-meter logs for a data center stopped in November; Gemini 3.1 Pro often missed the issue and sometimes judged the logs complete. |
| Consumer goods (+33 points) | A product launch assessment asked which capacity expansion to fund. The author says Argon found that existing inventory could cover the early capacity shortfall, so a smaller expansion would support an earlier launch and faster break-even. Gemini 3.1 Pro tended to recommend a larger expansion and a later timeline. |
The table below transcribes the accuracy bar chart in the attached image. Values are listed Gemini 3.1 Pro → Gemini 4 Argon; the chart provides no sample counts, confidence intervals, or per-task raw scores.
| Chart grouping | Gemini 3.1 Pro | Gemini 4 Argon |
|---|---|---|
| Full Dataset | 71% | 85% |
| Public Sector (Industry Subset) | 76% | 94% |
| Life Sciences (Industry Subset) | 61% | 88% |
| Healthcare (Industry Subset) | 57% | 87% |
| Legal (Industry Subset) | 64% | 75% |
| Media & Entertainment (Industry Subset) | 65% | 79% |
The chart shows a 14-percentage-point difference on the full dataset, consistent with the post's overall claim. The five subsets are categories shown in the chart; the five task examples in the post are a separate set of items. Their category names and margins are not interchangeable.
This is a first-party account by a Box executive of an internal Box evaluation, not an independent third-party assessment. The attached chart adds overall results and accuracy for five industry subsets, but Box did not publish the dataset size, task sampling, evaluators, scoring rules, or error margins.
The post does not disclose Gemini 3.1 Pro's call path; whether the two models used the same complete agent configuration; prompts; model snapshots; reasoning levels; temperature; tool calls; number of runs; cost; or latency. Apart from the statement that Argon used Box Agent, other configuration details should not be inferred.
The five examples in the post illustrate differences observed by the author. The +N-point figure for each example is not the full-dataset or industry-subset accuracy shown in the chart, and no corresponding raw outputs are provided for independent review.
The results apply to the specific internal evaluation version and enterprise-document tasks described by Box. They do not establish how the models perform on general knowledge, all enterprise workflows, or in other products. The professional-domain examples are records of evaluation results, not medical or financial advice.
What can be checked is limited to Aaron Levie's post text, its 2026-10-13 UTC timestamp, and the chart attached to the post. Reproducing the evaluation would require the Box AI V18 task set and data, scoring criteria, complete agent configuration, call parameters for both models, and records of repeated runs. The post does not make these materials public.
Gemini 4 Argon