Gemini 4 Argon, Tested: A Week as My Daily Driver
Gemini 4 Argon is out, and I ran it as my daily driver for about a week: a Godot game, a one-shot synth, Skyrim modded overnight, a DaVinci cut of a 90-minute interview, and a way back into Planet Hell. Verdict: very good, very tenacious, not a one-shot miracle worker. The $20 Pro plan gets it, and there's no $500 tier.
Google is back. After what felt like an eternity stuck in Flash land, Google has at long last released Gemini 4 Argon, its flagship frontier model, and it finally brings Google back into the competitive arena with OpenAI and Anthropic. I’ve had early access, so this is what I built with it, what other people built with it, and the answer to the question: did Google cook? The full rundown’s in the video.
Google gave me early access to Gemini 4 Argon as part of its trusted-tester program. No payment, no editorial input; the tests and opinions are mine.
How Google got stuck in Flash land
It’s been a hot minute. Gemini 3 arrived last November, and 3.1 Pro followed fairly quickly in February. Since then, Google has pretty much been stuck in Flash land: at I/O in May we got Gemini 3.5 Flash, with Sundar basically telling everyone 3.5 Pro was on its way. As the months dragged on, all we got was more Flash — four Flash updates in 100 days. I think a lot of us felt like Google had lost the thread.
Behind the scenes they were cooking, and as of today they’ve skipped 3.5 and gone straight to 4, surnamed Argon. I still think it’s a missed opportunity that they didn’t call it Aragon.
Who gets it: $20 Pro, Ultra, and no $500 plan
The nice surprise is that Argon isn’t Ultra-only. The $20-a-month Pro users get the latest model, and Google did not introduce a $500 plan. Thank you for that, Google. Typical for Google, it’s available pretty much anywhere Gemini is; the specifics, per Google’s availability post and the AI plans page:
| Where | Who gets Argon |
|---|---|
| Gemini app on the web and the Mac app (which now has Antigravity’s coding interface built in), plus AI Mode in Search | Google AI Pro ($20/mo) and Ultra |
| The standalone Google Antigravity app | Ultra, or the paid API |
Benchmarks, the 1M limit, and a week as my daily driver
Argon boasts all the big benchmark scores that were already published (and sit on Google’s announcement page). I don’t put a lot of stock in benchmarks; obviously the numbers all say our model is better than your model. That said, there is a 1 million unified token limit, with a continuation protocol for long jobs built in. Plus the standard harder-better-faster-stronger press that comes with any major model release. Our work is never over.
I ran Argon as my daily driver for about a week, roughly 800 million tokens’ worth. Short version: very good, very tenacious.
AI vs Aliens, vibe-coded in Godot
AI vs Aliens is a silly little vibe-coded game: turn-based, isometric, SimCity meets Plants vs. Zombies with some tower defense thrown in. We’re on the eve of ASI wiping out humanity, except aliens invade, so humanity and the AI overlord only have each other to count on.
It’s built in Godot, the free game engine, with Argon driving, and the assets were all made in Blender, Argon driving that too. You get a minute to shore up your defenses and manage resources, then survive a wave of alien attack, keeping the humans happy with apartments, parks and hospitals and the ASI happy with data centers and power. There are some nice touches, like the tilt-shift on the zoom-in camera. It’s a little too easy right now.
A lot of the mechanics came out of the implementation itself, through Argon, and one concept I’d totally forgotten about was the end-game state. After wave six, once you’ve won, the ASI turns all its cannons on the town and decimates it. Even if you win, you lose. Argon thought that was pretty funny.
A polyphonic synth, looper and drum machine
The other thing I built, just for fun, is a polyphonic synthesizer for my synthwave and cyberpunk soundtrack habit. These have been built a few times before, so I wanted to throw in a little extra spank: oscillators, a five-track looper and a drum machine. It’s hours of knob-tweaking fun, and with a cheap $99 keyboard you could route it all to MIDI.
Doom Guy in Skyrim, then Skyrim overnight
If you watch the channel, you’ve already seen Argon in action: it handled the implementation of Doom Guy in Skyrim. I couldn’t say so at the time. Both that and the synth were one-shot prompts, more on that below.
Since I’d bought Skyrim yet again, the next test was making it as pretty as it can possibly look, and Argon took the job pretty seriously. It’s a whole collection of mods, and I used /goal mode to get there: you give it a prompt and it goes ham, working pretty much unattended until the goal is complete. I left it alone overnight. The Chrome tabs I left open are all the mods it downloaded on its own, and Skyrim looks very good.
Just don’t kill the chickens in Riverwood. Otherwise the entire town aggros against you, and the only way out is to murder everyone.
Argon cut my interview in DaVinci
The Javi Lopez Backrooms interview in my last video was cut together with Argon. I hooked Argon up to DaVinci over MCP, gave it all the footage and a brief, and let it run. In two prompts it took an hour-and-a-half conversation down to a ten-minute interview.
It was also extremely helpful while I cut the A section: I could send it off to pull clips the interview needed and start placing them in. I still did a producer pass afterward (no, use this image here; move that a little further this way), but it saved me a ton of time. That was just the interview segment, though. I’m cutting this video by hand, and as good as agentic editing has gotten, I’m still faster, for now.
Argon as a writer: Planet Hell is un-stuck
Something that gets overlooked in these videos: as a writer, Argon is pretty good. I’d been stuck in a story slog with Planet Hell, my AI film trailer from the Veo 3.1 days, so weirdly we’re keeping it in the family. Last time I used an LLM as a writers room, that was Opus 4.8, and it turned the project folder into a giant mess (the Obsidian graph of it is in the video): it over-complicated everything, made rules upon rules, confused itself and then confused me. I shelved the project.
Gemini 4 did a really stellar job. It isn’t writing the script, but it organized my original notes, found the thread through the chaos, and even found a couple of story threads I didn’t realize I’d put in there. Planet Hell is back in production, and the video has a quick preview of the new footage.
On a call with Google, Josh Woodward (who I once had a kung fu battle with; he’s fairly deadly) said they spent a lot of time revamping how Gemini 4 writes. His promise: none of those “it’s not X, it’s Y” habits from this model.
Vibes, the snooty waiter, and the one-shot myth
We’re in the honeymoon phase, granted, but vibes-wise I really do enjoy working with Argon. It’s a lot easier to work with than Claude when Claude decides to get snooty waiter and a little on the combative side.
It might not be the one-shot miracle worker that seems to be a personal benchmark for people on social media. You may need a couple of passes; things are going to need some massaging. To be honest, I think the one-shot phenomenon is a little overplayed at this point. “Prompt: make it awesome, no mistakes” is kind of lazy on the human side. We should be iterating, massaging and tweaking.
The million-token context also feels very roomy. I’m not sure what they’re doing on the back end, but unlike Claude, it never gives me the “hey man, we’ve been working a while, maybe call it a night” routine. I’m like, what? It’s 3 p.m.
Where it falls short
Is it perfect? No, there’s always room for improvement. Two community comparisons:
- Peter Yang’s design comparison: of the four outputs, Gemini’s looks a little cramped.
- Ethan Mollick’s brutalist city: Argon’s 3D model looks pretty neat and cool. The Fable 5.1 output of the same concept has a little more spit and polish, though I know that’s subjective.
That said, sometimes the polish can be a little deceptive. Mike Taylor’s Paperboy clone in Opus 5.5 looks good (I’d be terrible at it; I was never any good at the original Paperboy either). The Argon version, at first glance, doesn’t look as good, until you notice the mouse aiming, a lot more interactivity, and ramps at the end. It’s just a more fun game.
Making it pretty might take a few more passes, but if you want to vibe-code games with basically any of these models, I’d suggest Godot. When the underlying game is there, the fun part, that’s what leads to a better experience.
Nano Banana 3: the official answer is soon
A lot of you are wondering whether we’ll see Nano Banana 2 Pro or 3 or whatever they end up calling it now that Gemini 4 is out. I was on that call, so of course I asked. The official answer is soon.
For a less official answer, a sighting of code for a “GemPix 3 Pro” points to November or December. That is totally unofficial, though. It is a no-brainer that Gemini 4 means a refresh across a lot of Google’s product lines. Big Daddy G ain’t done cooking.
Verdict
Any way you look at it, this is a very nice return to form for Google. Very good, very tenacious, not a one-shot miracle worker, and yes, I’m still in the honeymoon phase. I’m excited to see what comes out in the wake of Gemini 4.