We Tested Gemini 4 Argon For A Week! It's Surprisingly Good

Summary

The presenter reports that Gemini 4 Argon performed better in real-world testing than expected. After seeing its strong benchmark results, they initially suspected those results might be misleading.

What stood out

• Clear, well-structured writing: In a literary analysis test about Annie Dillard’s For the Time Being, Gemini organized a complex answer into numbered sections and developed a coherent explanation without making it feel overwhelming. In this comparison, the presenter found it more detailed and sequential than Claude, and more expansive than Astra.
• Explaining significance: In a job-description test, Gemini explained not only what a role involved but also why it mattered to the company’s strategy, even without being prompted. Its use of headings, bold text, and numbering made the response easy to follow.
• A useful role for writing and reading: The presenter sees Gemini as particularly helpful for unpacking complex information and explaining it clearly to readers, clients, or users. They sometimes use it for that work, then move the output to Claude for more creative projects.

Limitations

• Coding task incomplete: Gemini could not finish the presenter’s Senior Engineer test, which involved rewriting a codebase from scratch. The presenter was unable to get it through the full task and therefore did not assign it a rating.
• Poor development environment: The model had to be used through Anti-Gravity, which the presenter criticized for lacking basic workflow features—especially a classifier to automatically approve certain actions. Repeatedly clicking “always allow” made basic work frustrating. The presenter would rather access Gemini through the command line or another agent, though that could limit more advanced use.

Overall assessment

The presenter sees the release as a sign that Google could become a stronger competitor again. They don’t expect Gemini to replace established tools for people committed to the Codex or Claude ecosystems, but recommend trying it for tasks centered on writing, reading, and explaining complex material. They also value having another major model provider in the ecosystem.

Review and sources

Every says its team tested the model for several weeks across writing, reading, design, and programming. A fuller team assessment and collected test results are available at Every.to.

The main speaker is an unnamed presenter from Every, drawing on their own tests and the Every team’s broader evaluation. Other models referenced for comparison include Claude, Opus 5.5, and Astra.