On October 13, Box CEO Aaron Levie posted that Gemini 4 Argon beat Gemini 3.1 Pro by 14 percentage points on Box’s internal eval, after a week of testing.

I believe the number. I also would not move a production workload on it, and the reasons are specific enough to be useful.

What the 14 points are

Box published the details in a post by Aditi Tuli, a product manager on Box AI. Each task hands an agent a bundle of enterprise files (spreadsheets, PDFs, slides, images) and asks for the finished work product. Outputs are graded against weighted rubrics, and tasks run more than once.

The headline is 85% against 71%. Underneath, the gain is uneven. Public sector went from 76% to 94%, life sciences from 61% to 88%, healthcare from 57% to 87%, media from 65% to 79%, and legal from 64% to 75%. That is a spread of 11 to 30 points. By type of work, the biggest jump was data analysis at 23 points.

Here is what the post leaves out. It does not give the number of tasks, who or what does the grading, how each model was configured, or what the run cost. It says Argon has a tighter worst case on run time and gives no figure. And it measures Argon against Gemini 3.1 Pro. If your production system runs on a different vendor’s model, the 14 tells you nothing about the switch you are considering.

None of that makes Box wrong. It is Box’s eval on Box’s work. Levie’s thread even shows why it is a good one. Gemini 3.1 Pro often pulled the wrong rows from a spreadsheet, once reporting retention above 100%, and sometimes called a meter log complete when it stopped in November. Those are concrete ways to fail, and they tell you what to look for in your own files.

Price the pilot at $4 and $20

Google lists Argon at $2 per million input tokens and $10 per million output tokens through November 30, 2026. After that it is $4 and $20, per Google Cloud’s announcement. Same usage, double the bill.

A pilot costed at the launch rate hands your CFO a number that expires. Run the arithmetic at $4 and $20 every time.

A test you can run on your own documents

I wrote earlier about why price per token is not cost per job and why your evals are an asset that compounds. For a provider decision, the short version has five parts.

Start with the golden set. Take a few dozen tasks your people have already done, such as a spreadsheet analysis, a contract summary, a two-report reconciliation. The person who does that work writes the correct answer before any model runs. Include tasks where the right answer is “the data is incomplete,” because that is where Box’s older model slipped.

Then hold everything equal. Same prompts, same documents, same settings for both models. If you tune a prompt for one, tune it for the other or write down the difference. Run each task several times. In the study behind that earlier post, one model on one query varied in cost by up to 9.7 times between runs, so a single run is an anecdote.

Third, compute cost per correct answer: total spend, retries included, divided by correct answers. A model that costs more per token and finishes the job on the first try can come out cheaper.

Fourth, measure latency, the average and the worst case. Your users wait for the worst case.

Last, log how each model fails. Wrong rows, silent omissions, a job that quits halfway. Two models with the same score can fail in very different ways, and a failure you can see costs less than one that slips past.

Weight the errors by what they cost you. A wrong funding-gap calculation and a clumsy sentence in a summary are not the same score.

Write the switch rule first

Before the first run, set the threshold. How much must cost per correct answer improve, and how many serious errors will you accept? Set it after you see the numbers, and you are justifying a decision you already made.

Switching also has its own price: tuned prompts, a security and compliance review, a contract. When I read the GPT-6 Astra system card as a buyer, it was missing the independent sign-off a procurement file needs. That review applies when you change vendors too.

If Argon wins on your documents, switch. If it ties, stay where you are and test again once the promotional price ends.

If you want help building that golden set, write to us with the three document types that move the most money in your operation, and we will start there.