Level 3 · The server room
Field reports
When a major model comes out, I put it to work in my own studio and write up what I find, beside the benchmarks and what other people made with it in its first days.
How every report is weighed
A comparison only counts when every model worked in an equally good setup. Otherwise you’re measuring the workshop, not the model.
So each report says what every piece of evidence can and can’t show, and marks a result from an uneven setup as unestablished until it’s run again on even terms.
Gemini 4 ArgonGoogle DeepMind · out 13 October 2026
Field report 01 · 16 October 2026
A real frontier contender. Not the new leader.
106 benchmark results · 353 posts and videos · 21 graded jobs and four tests from my own studio
Read the report →