Paul J Lipsky's Gemini 4 Argon review is most useful where it goes beyond scores: he builds two games and a wearable-controlled prototype, asks Argon to inspect code from another model, writes with it, and gives it a recorded talking-head clip to edit through Antigravity. The impressive part is range. The useful part is that he also names the friction: game visuals needed more direction, the video still had errors, and Antigravity felt less smooth to him than his other coding tools.
Paul's week of early access makes Argon look versatile for prototyping, reviewing code, and writing. It is not a controlled model comparison or a security audit. He still prefers Opus 5.5 for some coding and for his AI video-editing workflow. Google's published benchmarks and rollout details help explain the model's ambition, but the purchase test is whether it completes your job with acceptable corrections, time, and cost.
Watch Paul Test Gemini 4 Argon
Credit: The examples and judgments below come from Paul J Lipsky's video. The Hyperagent segment at 03:38 is sponsored and is separate from his Argon tests. Product access and technical limits are checked against Google's linked pages.
Six Tests, With the Actual Limits
1. Two playable games
What Paul shows
In Antigravity, a paintball capture-the-flag game takes an initial build prompt plus a second prompt for a difficulty slider. A space-fighter game with enemies and a boss takes one main prompt.
What to check
Paul says the mechanics worked, but his first Argon games did not look as good as he wanted. He had to describe the visual direction more explicitly than with Opus 5.5. These are prototypes, not tested production games; the space-game reference also calls for original assets and design before publication.
2. A headband-controlled rocket
What Paul shows
Argon helps connect his meditation headband over Bluetooth to a small rocket game. He can move the rocket during his demonstration.
What to check
He says he has to exaggerate his expression, including raising his eyebrows. The clip shows a playable input experiment, not a validated measure of concentration, calmness, or direct thought control.
3. Bugs in another model's app
What Paul shows
He gives Argon the code for an Apple Watch-to-GrokBot app he built with Opus 5.5. Argon flags several issues, including a possible webhook-exposure path; Paul says he patched it.
What to check
The video does not publish a reproducible exploit, diff, or independent security review. Treat the finding as Paul's reported result. For your own app, verify the path, rotate any exposed secret if needed, test the fix, and keep private code inside approved tools.
4. Document answers and a Spark deck
What Paul shows
In the Gemini app he reports more direct answers and better adherence to supplied documents. In Gemini Spark he researches Argon and assembles a presentation.
What to check
These are his impressions from early use, not a measured hallucination study. He also says attractive slides are possible with smaller models, so this deck alone is not a reason to pay for Argon. Check every source claim and slide before sharing.
5. Long-form writing
What Paul shows
He asks for a scary story long enough to read in roughly 20 to 30 minutes and says it held his attention. He also likes Argon for scripts and technical writing.
What to check
Writing quality is subjective, and technical or legal prose needs domain review. Paul's story link redirected to a Gemini “Link doesn't exist” message when checked on 14 October 2026, so we could not read the full story independently.
6. Talking-head edit and motion graphics
What Paul shows
Argon writes a short script, Paul records himself reading it, then asks Antigravity to edit the footage and add motion graphics. He likes the result despite a few mistakes.
What to check
Google lists Argon's API output as text. This is a code-and-tool editing workflow, not native video generation by the model. Review captions, timing, graphics, sound, and rights before using a finished cut.
Where Paul Still Prefers Opus 5.5
Paul calls Argon one of the strongest models he has tried for video editing, but says Opus 5.5 remains his first choice there. He also finds that Opus reaches a pleasing game design with less visual prompting and that Antigravity is a weaker everyday harness for him than the ChatGPT app and Codex. Those are valuable practitioner observations, not a matched test: the same prompt, tools, revision budget, and scoring rubric were not held constant. The result may change with a different project or a better harness.
Google's launch benchmarks show strong performance across many of its published evaluations. Paul notes that Google reports Argon leading 13 of 19 listed benchmarks and tying one more. That is a company-reported snapshot, not evidence that Argon wins every coding or creative job. His own tests show a more useful split: strong breadth and writing, with extra visual direction and quality control still required.
Where Argon Is Available
Google's Gemini app update puts Argon in the rollout for Google AI Pro and Ultra subscribers. Its developer update covers the paid Gemini API and AI Studio and describes Antigravity access for Google AI Ultra. Availability in one surface does not guarantee it in another. As with any rollout, check the model picker and account entitlement before planning a client project around it.
For developers, the model card lists image, video, audio, PDF, and text inputs, but text output. Google separately documents the gemini-4-argon model ID and current paid API pricing. A Google AI Pro app plan is not an API allowance. This matters when estimating the cost of a service built on Argon rather than simply trying it in the app.
Jump to Paul's Video
| Time | Segment | Time | Segment |
|---|---|---|---|
| 00:00 | Introduction | 00:18 | Model and benchmarks |
| 01:12 | Games in Antigravity | 03:38 | Hyperagent sponsor segment |
| 04:55 | Finding app bugs | 05:41 | Gemini app |
| 06:26 | Spark and writing | 07:58 | Editing and graphics |
| 09:21 | Verdict and availability |
Business Idea: A Small-App Quality Clinic
Paul's cross-model bug check suggests a focused service: review a small app's real workflows, reproduce a few meaningful issues, and deliver a clear fix-and-verification package. The product is the verified improvement, not an impressive-sounding model diagnosis.
Test an app quality clinic
Scope one authorized app and prove one useful improvement.
Sources and Caveats
Creator source: Paul J Lipsky's review and the story URL he shared (unavailable when checked). Official sources: Google's announcement, app availability, developer access, and model capabilities. Paul's tests are one creator's early-access examples; his qualitative comparisons, reported bug, and writing reaction have not been independently replicated for this article.