TL;DR
The short version
After a week with Gemini 4 Argon, Dan Shipper says Google's model is surprisingly strong at clear, structured writing — and weak at finishing Every's hardest coding rewrite. The model may be back; the IDE is still in the way.
On Every / AI & I, Shipper compares Argon to Claude Opus 5.5 and Astra on literary and business writing, then reports a hard stop on Every's senior-engineer codebase rewrite and constant friction inside Anti-Gravity.
Built on We Tested Gemini 4 Argon For A Week! (13 Oct 2026) from Every, with the full team vibe check referenced at every.to.
Structured writing is the surprise
Shipper's standard literary probe is a hard question about a book he knows well — Annie Dillard's For the Time Being — with no single right answer. Argon answered in numbered sections that build. Claude Opus 5.5 tended to jumble; Astra was tighter but less detailed.
Gemini does a really, really good job of giving you a long, detailed answer that's not overwhelming.
Dan Shipper on Every / AI & I
The same pattern showed up on a job-description task: Argon stated what the role is and why it matters to company strategy without being told to.
The coding bar it did not clear
Every's senior-engineer benchmark asks a model to rewrite a codebase from scratch end to end. Argon could not finish. Shipper recorded no score. That is a hard stop on the task that matters most to teams living in agentic coding tools.
Anti-Gravity is the real friction
Google ships Argon through Anti-Gravity. Shipper found basic agent loops painful: no classifier to auto-allow routine actions, so constant "always allow" clicks. His workaround is to call Argon from another agent's command line instead of living inside Anti-Gravity.
This is the first release where I'm like, oh, they actually could be back.
Dan Shipper on Every / AI & I
How to try it without switching stacks
If you are deep in Codex or Claude Code, Argon will not become your daily driver from this vibe check. If you have reading and explanation tasks that need long, non-overwhelming structure, it is worth a try — preferably from a CLI you already trust.
One team's week is not a market ranking. Run your own blind tasks before you switch. The full Every team write-up is on every.to.
