Gemini 4 Argon (Fully Tested & Thoughts): It's Just Fake? I Tested It on 6 Hard Apps

In this video, I'll look at Google's new Gemini 4 Argon, starting with its pricing and the launch benchmarks, and why some of those charts aren't apples to apples. Then I'll go through what independent testers like Artificial Analysis, Vals, Zapier and Surge actually measured, where Argon leads on office work but is mixed on coding. After that, I run it on my KingBench 4 tests through Devin: an emulator, a Docker alternative, a 3D room designer, a QMK keyboard app, Pokémon Red and a Fitbit desktop app. Finally, I show the pattern I noticed, where it fakes the hidden parts of an app, and give my verdict on whether most people should use it. -- Key Takeaways: 💸 Gemini 4 Argon launches at $2 / $10 per million input / output tokens, and the price doubles to $4 / $20 after November. 📊 Google's launch charts have fine print, like a best-of-three-runs computer-use score and six times more checker time on a science test. 🧠 Artificial Analysis found Argon makes stuff up only 15% of the time when it doesn't know, versus 51% for Astra and 54% for Sol. ✅ Argon is number one on the Vals Index and leads Zapier's AutomationBench, but it still fails about half of those tasks. 💻 On coding it's mixed: an app-building tie at a much lower cost, but it trails on Terminal-Bench and writes messier code. 🎮 On KingBench 4 it made the best emulator and the best Pokémon game of any model I've tested. ⚠️ It also made the worst keyboard app, which never actually talks to your keyboard. 🔍 The pattern: the visible parts look great, but hidden parts like the Fitbit sync and the keyboard connection are faked and labelled as real. 👍 Argon scores 32.6 out of 60, fourth behind Opus 5.5, Astra and Sol, and basically tied with Haiku 5.5. 🎯 Verdict: a real frontier model and a good comeback for Google, but not the Opus killer the charts suggest, so most people shouldn't switch to it.