Google Goes Frontier: Gemini 4 Argon, Tested

0:00"Google's Gemini 4 Argon is here, and yes, Google has caught up." Matthew Berman has had access for about a week, and this is his hands-on verdict on the model that finally puts Google back at the frontier alongside OpenAI, Anthropic, and Meta.

Google goes frontier

0:12His first test: "the most incredible hyperrealistic 3D animation of a cyberpunk city" in Three.js, running at 60fps with neon colors and a flying car. The result is genuinely impressive — tons of 3D buildings, animation, and an interactive control panel (color presets, neon bloom, ultra quality, camera director). One behavioral quirk he's noticed: 1:19without specific guidance (or a /goal), Argon tends to want to spend around 30 minutes on a task by default.

The 3D tests

2:19A 3D Rube Goldberg machine in Three.js: fully controllable, with realistic physics, multiple themed levels (Grand Canyon, Las Vegas, clockwork lab, Kyoto temple), auto-follow camera, and a pinball-style chain reaction. He calls the physics "really good," with minor clipping. 7:45An Unreal Engine ocean scene with a boat: high-detail water and good waves, but he couldn't go underwater — a contrast with other recent models that can. His honest read: Argon is "much better than Muse Wonder at 3D, still not quite as good as Opus and Astra."

Argon vs. Muse Wonder on 3D

1:31He uses the opportunity to compare against Meta's Muse Wonder (released the prior week, covered in our Muse Wonder deep dive). Muse Wonder is "an absolute frontier model, but it's not good at 3D — not 3JS, not Unreal Engine, not Blender." Berman attributes this to a specific admission: Meta's head of AI, Alexander Wang, reportedly said their team "either forgot or didn't RL the model to be really good at 3D," calling it a mistake in retrospect. Neither model is weak overall — just weaker on 3D than Opus and Astra, which he calls "groundbreaking" on that axis.

The Excel clone

8:15He also gave Argon the now-standard reverse-engineering test: clone Excel. With no loop set up (only ~20 minutes of runtime), it opened Excel on his machine to inspect it, then produced a working browser clone — =1+2, =sum, drag-to-fill, and row highlighting all function. 9:15His takeaway is blunt: "software is basically going to cost nothing in the future. If the only moat you have is the lines of code you write, you're going to be in trouble."

Benchmarks

4:35First, the sponsor-provided numbers: Box (this video's partner) reports a 14-point jump on their knowledge-work benchmark suite — 71% to 85% from Gemini 3.1 Pro to Argon — across spreadsheet evaluation, PDF extraction, presentations, and images. 10:44Then the model-vs-model chart (Argon vs. Astra, Fable 5.1, Opus 5.5, plus a second chart he rebuilt to add Muse Wonder Full and GPT-6.1 Sol): DeepSWE 77.9% (second, behind Muse Wonder's 79); VibeCode Bench 91.9% (number one); Terminal-Bench 57 (behind Opus 5.5 at 66.4, and behind Muse and Astra); OSWorld 69% (vs GPT-6.1 Sol's 71.4); Agent's Last Exam 39 (vs Opus 38). 12:00His verdict: "Google is definitely at the frontier," and for the first time all four big labs — Google, Meta, OpenAI, Anthropic — have competing frontier models. "When there's lots of competition, we win."

The rollout

9:36One oddity he flags: Google announced Argon a couple weeks before granting public access — benchmarks first, model later — with early testers and security researchers getting it quietly before the public release. Typically the whole thing drops at once. That matches Google's own timeline: announced September 30, 2026 (via its Fairwind cyber-defender program), broadly available October 13, 2026.

Claims checked

Sources & further reading