Matt Wolfe's Gemini 4 Argon test is a useful companion to Google's launch materials. His one-shot game plays, his physics playground responds to controls, and a generated STL makes it into a physical print. But a nice prototype is not a finished product: the game's styling trails his favorite rivals, the print's side profile is weaker, and two Gemini Tasks do not clearly finish. The question is not whether Argon can make impressive things. It is which jobs it completes reliably at a price you can justify.
Gemini 4 Argon looks competitive for coding and interactive prototypes in Matt's early test, but this video is not a controlled head-to-head benchmark against Claude or ChatGPT. Google's million-token output ceiling and benchmark results are notable; the stalled Tasks are a reminder to measure completion, not just intermediate artifacts. As of October 13, Google has expanded access across the Gemini app and developer surfaces, with separate paid API pricing.
Watch Matt Wolfe Test Gemini 4 Argon
Credit: The demonstrations, reactions, and comparative opinions are Matt Wolfe's. Availability and prices below come from Google Cloud's October 13 update and the developer pricing page; the independent benchmark context comes from Artificial Analysis.
Where You Can Use It, and What It Costs
Google announced Argon on September 30, initially giving selected cyber defenders access. The October 13 rollout expands it to Gemini for business on web, desktop, and mobile, and to developers through the Gemini platform API (gemini-4-argon), Google AI Studio, and Google Antigravity. The Gemini app release notes describe a global rollout for AI Pro and Ultra subscribers. Matt demonstrates Argon in Antigravity and in his Ultra Gemini app account; availability may still depend on your plan and rollout state.
Do not confuse an app subscription with API credit. Google's developer API pricing page says Argon is unavailable on the free API tier or through a Google AI Pro plan. Standard API pricing is $2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20 from December 1; output pricing includes thinking tokens. Implicitly cached input is $0.10 per million tokens during the introduction, then $0.20. The page lists other service classes, so check the billing mode before estimating a project. Matt's early-access demo account did not provide reliable per-demo token or cost figures for his SVG test.
A Million Tokens of Output Is a Ceiling, Not a Deliverable
Google's announcement and API documentation describe output generation up to one million tokens. That is an output limit, distinct from the input context window. It does not mean an ordinary request will produce that much, or that a very long result will automatically be coherent, verifiable, or affordable.
Google reports strong results on professional-work and agent evaluations, including 77.9% on DeepSWE 1.1 and 51.3% on AutomationBench in its launch materials. These are benchmark results under their specified methods, not measured success rates for your application. Artificial Analysis independently scores Argon at 53 on its Intelligence Index in its September 30 assessment, alongside GPT-6 Astra's maximum-effort score and above GPT-6.1 Sol's maximum-effort score in that snapshot. An index point is not a percentage. Its cost-per-task estimate uses a defined benchmark workload and introductory prices; both comparison scores and prices can change.
What the Hands-On Tests Actually Show
1. SVG portrait
Visible result
Matt asks for a Gary Busey portrait in SVG and judges the image among his strongest examples so far.
Practical limit
He had not yet placed this result on his leaderboard and could not report representative token usage or cost from the demo account. Visual taste is not a controlled ranking.
2. Playable game
Visible result
His MegaBonk-inspired one-shot build has character and stage selection, camera movement, automatic attacks and leveling, jumping, and sliding. He finds the mechanics solid on first play.
Practical limit
He prefers the visual design of his Opus 5.5 and GPT-6 Astra examples. Those are creator impressions across builds, not a prompt-, tool-, and revision-matched comparison. A business game also needs original assets and rights review.
3. Physics playground
Visible result
The interactive cannon simulator exposes Earth, Mars, and Moon gravity presets, angle, speed, wind, and target controls. Changes update the projected trajectory in real time.
Practical limit
The video shows responsive interaction, not a numerical validation of the physics. Test trajectories against known equations before using it to teach or make decisions.
4. Scheduling constraints
Visible result
Given two people, task constraints, and a lunch interruption, Argon returns a schedule, a 2 p.m. earliest finish, and an explanation that Matt considers good.
Practical limit
The full scheduling instance is not reproduced in the video description, so this post does not independently prove optimality. Recheck each constraint and lower-bound argument before relying on the schedule.
5. Plan comparison
Visible result
Argon presents a polished side-by-side answer to whether ChatGPT Plus, Claude Pro, and Google AI Pro cover use of their models in a separate app.
Practical limit
A good comparison layout is not proof that plan entitlements are current. In particular, Google's Argon API page treats AI Pro app access and developer API billing separately.
6. Printable 3D bust
Visible result
From a reference image, Argon generates an STL that Matt opens in Bambu Studio and physically prints. He likes the front view and shows the finished object.
Practical limit
The side profile looks weaker to him. One successful print proves this file was usable in his setup, not that all generated meshes are watertight, accurate, or ready for commercial use of a person's likeness.
The Tasks Stalled; the Robot Was Google's Demo
At 15:00, Matt moves from chat output to Gemini Tasks. He asks one task for AI news research and another for slides. After roughly an hour, the research task has not returned a finished answer; the slide task has produced several slides but still appears to be working. His account was using early-access features. The video cannot establish whether the bottleneck was Argon itself, the Tasks product, or the environment. For a work process, count a job as complete only when the requested artifact is usable and the task state is resolved.
At 16:30, he also plays Google-supplied footage of a robot learning to throw at a cornhole board. This is not a robotics test Matt ran himself, and the short clip does not establish a general real-world success rate. It is a separate demonstration from his coding and Gemini-app tests.
Has Google Caught Up?
In Matt's view, Argon belongs in the frontier-model conversation, especially for coding and fast interactive prototypes. His video supports a narrower conclusion than “best model”: one good SVG, one playable but less polished game, a convincing interactive simulation, a plausible schedule, a printed bust, and two unfinished Tasks. The independent index puts Argon near other leading models in its release snapshot, while Google's specialized benchmark results show strength on their measured tasks. Neither settles the question for every developer or business.
For a real purchase decision, pick one authorized task, compare the same prompt and tools across models you can access, and score accepted result, correction time, task completion, and total cost. If the job can cause a purchase, send a message, change a system, or influence a high-stakes decision, keep human review in the loop.
Jump to the Video
| Time | Segment | Time | Segment |
|---|---|---|---|
| 00:00 | Gemini 4 Argon arrives | 01:47 | Benchmarks and output limit |
| 04:32 | Antigravity and Gemini app | 06:54 | Access and pricing |
| 08:21 | SVG, game, and physics tests | 11:38 | Scheduling and research |
| 13:34 | 3D print | 15:00 | Gemini Tasks |
| 16:30 | Google robot demo | 17:35 | Matt's verdict |
Business Idea: Interactive Training for One Repeated Task
The physics playground suggests a practical service: turn one hard-to-explain business procedure into a small interactive training tool, then verify the behavior with a subject-matter expert. Sell the tested outcome, not the model name or the size of its output window.
Prototype a training simulator
Start with one recurring problem and a seven-day validation test.
Sources and Limits
Creator source: Matt Wolfe's video and chapter-linked demonstrations. Official Google sources: September launch and evaluations, October 13 Cloud availability, Gemini app release notes, and API prices and exclusions. Independent context: Artificial Analysis. The benchmark snapshot predates today's wider rollout; access, quotas, prices, and rankings can change. The hands-on demos are not a controlled model bake-off.