We Tested Gemini 4 Argon For A Week! It’s Surprisingly Good
Every YouTube 56,429 views across 2 copies, as of 13 October 2026
Why it is here
Every’s video version (5:25) of its ‘Vibe Check’ after a week of early access, the review the timeline entry quotes (strong for writing and from-scratch builds, weak on Dan Shipper’s refactoring benchmark). Dan Shipper posted a 5:25 video on X at launch; it is listed as a copy on the basis of length and timing.
Description
Description written by Gemini from the videoGemini 3.8 Flash, 13 October 2026
Summary
Dan Shipper, co-founder and CEO of Every, reviews Google’s newly released Gemini 4 Argon after hands-on testing by the publication’s team. While initially skeptical of Google’s benchmark claims, he highlights that the model excels at structuring long-form writing and explaining complex conceptual context, despite failing their multi-file code rewriting benchmark and suffering from usability friction inside the Antigravity environment.
What is shown
- [00:09] A benchmark comparison table displaying Gemini 4 Argon against GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 across agentic coding, knowledge work, and science benchmarks.
- [00:31] A side-by-side text analysis test comparing Gemini 4 Argon and Claude Opus 5.5 interpreting recurring imagery in Annie Dillard’s book For The Time Being.
- [02:03] A job description output generated by Gemini 4 Argon (“Research Lead, Future of Coding”) demonstrating how the model explains the strategic thesis behind the role.
- [02:45] The Senior Engineer Benchmark scoreboard (
checks.every.to/senior-engineer-benchmark) testing models on full codebase rewrites from scratch. - [03:19] The settings modal in Google’s Antigravity developer environment displaying tool permissions and execution approval options.
- [03:48] Articles and media coverage examining Google’s AI market share and previous Gemini iterations.
Claims & numbers
- Benchmark scores displayed for Gemini 4 Argon: Vals Index (68.9%), AutomationBench (51.3%), Vals Finance Agent v2 (65.4%), Harvey’s Legal Agent Benchmark (19.6%), DeepSWE v1.1 (77.9%), FrontierSWE v2 (55.0%), Vibe Code Bench (91.9%), Terminal-bench 4.0 (57.4%), PostTrainBench (45.3%), Terminal-Bench Science 0.1 (57.6%), LABBench 2 (88.8%) [00:09].
- The presenter states that Gemini 4 Argon could not finish the task on Every’s Senior Engineer Benchmark (rewriting a codebase from scratch end-to-end) and therefore received no score [02:53].
- The presenter claims Antigravity lacks an automated classifier to pre-approve tool actions, forcing the user to repeatedly grant manual permissions [03:18].
- The presenter claims Google historically held the best price-to-performance-to-speed ratio among frontier providers before losing momentum over the past year [04:00].
Notable quotes
- [00:06] “And I was ready to hate this model. They released benchmarks a couple weeks ago, and they showed the model being the best at like every task.”
- [01:48] “Gemini does a really, really good job of giving you a long, detailed answer that’s not overwhelming.”
- [02:53] “And it just couldn’t finish the task. Like, couldn’t figure out how to get the model to actually do the thing end-to-end, so I don’t have a score.”
Assessment
An independent hands-on editorial review and “vibe check.” The presenter demonstrates authentic comparative prompt outputs and benchmark logs, clearly separating subjective workflow impressions from measurable coding test failures.
Described by gemini-3.8-flash on 2026-10-13 from the video’s audio and frames.