Gemini 4 Argon review: what the public tests show

This review collects what can be checked today: Google's own announcements, independent results from Artificial Analysis and Vals, and public tests people ran after Argon opened to paying users on October 13, 2026. Every number links to where it came from.

Where Argon stands on benchmarks

Argon leads on the benchmarks Google picked; on independent leaderboards it ties GPT-6 Astra and trails Claude Opus 5.5 by five points.

Sources: 9to5Google, Trending Topics and eesel. Nobody has reproduced Google's DeepSWE run independently yet.

Argon's clearest strength is that it makes things up less. On AA-Omniscience it gave a wrong answer instead of declining on 15% of questions, against 51% for GPT-6 Astra. The price is accuracy: it answered 50% correctly, 13 points behind Astra.

Its clearest habit is writing a lot. On the Artificial Analysis tasks it averaged 62,000 output tokens against Astra's 27,000, so a lower price per token is not a lower price per task: 1.99 USD a task at the launch price and 3.98 USD at the full price, against 3.26 USD for Astra and 5.98 USD for Opus 5.5.

3D, front end and vision

Argon builds 3D scenes in code quickly and cheaply, but in public side-by-side tests it does not produce the most detailed result, and it trails at rebuilding pages from screenshots.

• Google's own demos include pagoda and temple models in Blender, a desk racing game in Three.js, a ray-tracing engine in pure Python and a Gothic scene built from one image (Google).
• Same brief, two models. ohmyunicorn gave Argon and Opus 5.5 the same brief and 223 KB of notes to build a Three.js cartoon planet you can orbit and zoom. The author found Opus's planet the most detailed up close. Argon, running in Antigravity, checked its own page with screenshots and produced a cleaner build, but it also put a benchmark row on the page that does not exist.
• Rebuilding from images. In paddo.dev's vision runs (three each, scored 0 to 1), Argon scored 0.754 at turning screenshots into HTML, against 0.831 for Opus 5.5 and 0.836 for GPT-6 Astra. Drawing SVG from a text prompt it scored 0.614, ahead of Opus (0.503) and behind Astra (0.733).
• Long video. Google reports 91.7% on LVBench, its strongest published vision number, not yet reproduced (9to5Google).
• Front-end design. On launch day Bloomberg reported Google staff saying Argon does worse in practice than its benchmarks suggest, front-end design especially; Google called that "completely inaccurate" (ifeng, reporting Bloomberg).

Videos with narration that people credit to Argon come from the Gemini app around it. In Google's own Argon-made video, the voiceover came from Gemini 3.8 TTS and the music from Lyria 3.5.

Same task: time and cost

On single-prompt builds Argon is usually faster and cheaper. On real code with tests it was the most accurate model and also more expensive than Opus 5.5.

The production-code test is the more useful one: the tasks came from a real codebase and were graded by the tests that shipped with the original fixes. Argon passed most often and broke no test that already passed. It also wrote 1.37 million output tokens, 2.8 times as many as Opus, so it cost 1.26 times as much at the launch price. At December prices the same 20 runs would cost about 115 USD, roughly 2.5 times Opus.

The harness matters too. On the same change, Argon averaged 991 seconds in Claude Code and 1,740 seconds in Antigravity.

Robotics and physics

Every robotics and physics result so far is Google's own, with no failure rates and no independent reproduction. Google presents the robotics work as a step toward physical AGI, but positions Argon for coding, knowledge work, cybersecurity and science.

Sources: Google and TechEBlog.

The cornhole pipeline turns ordinary video into a simulation and then into robot motion. It does not let a model watch a video and drive a robot directly. Seven shots is a small sample, and Google has not said how much human setup each demo needed.

In science, Google lists three results: with Professor Matt Landreman, a broader class of counterexamples to Grad's 1967 conjecture on stellarator magnetic fields; five new serine esterase designs with catalytic activity confirmed in the lab; and quantum work including an arcsin circuit, verified in Lean, that cuts spacetime cost more than 100 times against a published baseline.

Who should skip it

• Pro subscribers who want Argon in Antigravity: it is Ultra only.
• Teams that pay per token for long production coding runs: December pricing likely makes it dearer per fix than Claude Opus 5.5.
• People whose work is mostly front-end design or rebuilding pages from screenshots: it trails Opus 5.5 and GPT-6 Astra there.
• Free and Google AI Plus users: they have no access.