How to test Gemini 4 Argon before you trust it

Gemini 4 Argon asks the practical question every team now faces: is the newest Google model good enough to replace the one in your workflow? A tester's week with it produced strong 3D demos and a usable Excel clone, plus lower terminal scores than three competing models.

Gemini 4 Argon benchmarks: where the model actually stands

Gemini 4 Argon is Google's newest flagship model, and the benchmarks its tester cites place it at the frontier in some areas and behind in others. The headline gain is a 14-point jump on knowledge-work tasks, while terminal control and photo-real 3D remain relative weak spots. Google had not published a public model card for this name at the time of writing, so treat every figure below as tester- or vendor-reported.

The model's rollout was unusual. According to the tester's account, Google announced the model and released benchmarks a couple of weeks before public access and let early testers and security researchers try it first. That matters for reading the numbers: the published scores were measured before later models were added to the same charts.

Benchmark results reported by the tester

Benchmark	Gemini 4 Argon	Best comparison cited	What it measures
Knowledge work (Box)	85%	71%, Gemini 3.1 Pro	Document and data extraction
Deep Sweep	77.9	79%, Meta Muse	Engineering sentiment index
Vibe Code Bench	91.9%	Roughly tied	Coding from prompts
Terminal Bench	57	66.4, Opus 5.5	Writing and running shell commands
OS World	69%	71.4%, GPT 6.1 Soul	Controlling a computer
Agent Exam	39	38, Opus	Agentic task completion

The pattern is consistent. Argon wins or ties on coding-from-prompts and general agent work, sits within a few points on desktop control, and loses ground on terminal execution, where the tester reports Opus 5.5 at 66.4 against Argon's 57. None of these results was independently reproduced for this article.

Knowledge work: the 14-point gain over Gemini 3.1 Pro

Gemini 4 Argon scored 85% on the total Box knowledge-work dataset, up from 71% for Gemini 3.1 Pro, a 14-point gain reported by the vendor behind the benchmark. Box, a cloud content-management company, published the comparison as a sponsor of the tester's video, so the result is vendor-reported rather than independent.

The suite covers tasks that match real office work: spreadsheet evaluation, PDF extraction, presentation structuring and image handling. The tester describes the same pattern across industry-specific subsets, with large jumps in legal, life-sciences and healthcare scenarios. The comparison uses a single prior model, so the 14 points measure progress within Google's own line, not Argon's lead over other vendors.

A 14-point rise is large enough to change deployment decisions. Teams running document pipelines on Gemini 3.1 Pro have a concrete reason to retest extraction accuracy, especially where sparse charts can make a few points look bigger than they are. The benchmark is a vendor instrument, and Box's own documentation defines the task mix.

Coding and agent tasks: a strong Vibe Code Bench score, a weaker terminal result

Gemini 4 Argon reached 91.9% on Vibe Code Bench, the top score in the tester's chart, but scored 57 on Terminal Bench against 66.4 for Opus 5.5. Vibe Code Bench measures whether a model can produce working code from a prompt; Terminal Bench measures whether it can write and run commands in a real shell. The gap between the two is the most useful signal for developers.

The split suggests a model that writes plausible code and less reliably operates the environment around it. A high Vibe Code Bench score helps with scaffolding, prototypes and single-file builds. A 57 on Terminal Bench means an agent that must install dependencies, read error output and recover from failed commands will need more supervision than one scoring in the mid-60s.

The tester also reports OS World at 69% against 71.4% for GPT 6.1 Soul, a near-tie on general computer control, and Agent Exam at 39 against 38 for Opus. Those results support a claim of parity on general agent work. They do not support a claim of leadership across agentic tasks, because terminal execution is part of that category and Argon trails there.

What the 3D demos show, and what they do not

Gemini 4 Argon produced a cyberpunk city, a Rube Goldberg machine, an Unreal Engine ocean scene and an Excel clone from short prompts, and the tester reports that the city ran at about 60 frames per second in the browser. The demos are the most striking part of the review and the weakest part of the evidence, because they are one person's prompting sessions rather than a repeatable test.

The tester's own assessment is mixed. The Rube Goldberg simulation showed visible clipping, and repeated prompts to make the scenes photo-realistic did not get there. The tester attributes part of that to the limits of rendering in the browser rather than to the model alone.

What each demo tested

1. The cyberpunk city was a browser 3D scene with a 60fps target, neon lighting, sky traffic and a camera director, built in roughly 30 minutes.
2. The Rube Goldberg machine was a controllable physics simulation with themed levels, including a Grand Canyon version and a Kyoto temple.
3. The boat scene ran in Unreal Engine and rendered ocean waves and a detailed hull, but the camera could not go underwater.
4. The Excel clone opened the real application to study its interface and reproduced formulas, row selection and basic layout in about 20 minutes.

The tester says Argon is clearly better at 3D than Meta's Muse Wonder, a model released a day earlier, but not as good as the models the tester calls Opus and Astra. That ranking rests on side-by-side demos, not on scored benchmarks, and should be read as a practitioner judgment.

Time, cost and the code-moat problem

Gemini 4 Argon tends to spend around 30 minutes on an open-ended build task, according to the tester, unless a time budget or goal loop is specified. That habit shaped every demo in the review. The city, the Rube Goldberg machine and the Fall Guys clone each consumed roughly half an hour of autonomous work before the tester intervened.

The tester's read on the Excel clone is blunt: if the only asset a software company has is its own source code, that asset is vulnerable. The argument goes beyond one model. A 20-minute browser spreadsheet with working formulas and a clean layout shows how quickly a surface-level product can be reproduced, which raises the value of data, distribution and services that cannot be copied from a prompt.

The Fall Guys clone supports the same point with a caveat. The tester reports that early builds flickered badly and that quality improved over the testing period, which could reflect model updates, accumulated prompt context or both. A reviewer cannot separate those causes from the outside.

How to evaluate Gemini 4 Argon for your own work

The right test for Gemini 4 Argon is a task you already measure, run against the model you use today. Published benchmarks tell you where to look; only your own evaluation tells you whether the model earns a place in a workflow. The tester's figures point to knowledge-work extraction and prompted coding as the strongest candidates and shell-heavy agents as the weakest.

1. Pick one document-heavy task, such as PDF field extraction, and score both models on the same 50 documents.
2. Pick one agent task that requires shell commands, and count how often the agent recovers from a failed command without help.
3. Run a fixed prompt with an explicit time budget, because the model may otherwise keep working for around 30 minutes.
4. Record the model version and access date, since the tester observed behavior changing during the review period.

Treat vendor benchmark pages as a filter for which tasks to test, not as proof of quality. The Box results are the clearest example: they come from the benchmark owner and were run before the newest competing models appeared on the same charts. For agentic code work, the developer documentation and release notes from each vendor are the place to confirm current model names and capabilities before you commit.

FAQ

• Is Gemini 4 Argon really a frontier model? The tester's benchmark charts place Gemini 4 Argon at or near the top on prompted coding and general agent tasks, and within a few points of the leaders on desktop control. It trails on terminal execution and photo-real 3D, so the frontier label holds for some workloads and not others.

• How much better is Gemini 4 Argon than Gemini 3.1 Pro? On the Box knowledge-work dataset the reported score rose from 71% to 85%, a 14-point gain. That figure compares Argon only with the earlier Google model and was reported by the benchmark's owner.

• What is Argon weak at? The review reports a Terminal Bench score of 57 against 66.4 for Opus 5.5, repeated failures to reach photo-realistic rendering, and visible clipping in 3D physics scenes. Those are the areas where the model trails its closest competitors.

• How long does Gemini 4 Argon work on a task? The tester observed a default of roughly 30 minutes on open-ended build prompts when no time budget or goal loop was set. Specifying a budget changed the behavior in testing.

Source video