Model test · October 14, 2026
Gemini 4 Argon vs four rivals
Google's new model against GPT-6.1 Sol, GPT-6 Astra, Claude Opus 5.5 and Claude Sonnet 5.5. Ten tests, three effort levels, the exact same prompts. Here is every number, every recording, and every file the models wrote, running live.
Final scoreboard · out of 10
Close at the top. Not close on price.
Each score is the average of ten tests at three effort levels. Cost is all 30 runs at API prices; time is wall-clock time, added up.
The two judges didn't agree
The GPT-6.1 Sol judge gave its highest visual marks to Sol; the Claude judge put Argon first. With only the Claude judge, Argon ranks first overall; with only the GPT judge, fourth. The formula, the average of both judges, was fixed before any result came in, so Sol stays on top. At the top, it is effectively a tie. That is exactly why the crowd vote below exists.
All 150 runs, one square each
Brighter means a higher score. Hover or tap a square for the details.
Playground · live code
Don't trust the recordings. Run them.
Every output here is the exact file a model wrote, running live in your browser. Put any two side by side, play the Snake games, or feed the parser and the calculator your own input and watch fifteen versions answer at once.
Each file runs in a locked frame: it can't touch this page, your cookies or the network. Nothing was edited; if a model's file has a bug, you'll see the bug.
Crowd judge · live
You be the judge.
Two recordings of the same prompt at the same effort, names hidden. Pick the better one. Then we show who made them and what the AI judges said. Once enough people vote, we'll know whether humans and machines see the same winner.
Keyboard: ← left, → right, Enter for the next one.
The crowd vs the AI judges
The recordings · 150 runs
Every run, the way the judges saw it.
Each model worked inside its own coding agent (Antigravity, Codex CLI, Claude Code) at low, medium and high effort, tools off. These are the screen recordings the judges scored. Click one for the details, or hide the names and pick your favourite.
Names are hidden and rows are shuffled. Click the result you like best.
Read the exact prompt
The other six tests
The snake games were played by a robot. The two functions met hidden test suites. The puzzles and trick questions had one right answer each.
Word for word
The ten prompts. Run them yourself.
Every model got exactly this text and nothing else. Try them on any model you like, then compare with the files above.
- Tools turned off; every model answered in one go.
- Claude Code's personal CLAUDE.md loading was switched off, and all 60 Claude runs were redone.
- Visual work: two blind judges from two companies, fixed rubric. Code: hidden tests and a robot player.
- Cost: tokens × each company's API price on October 14, 2026.
Your call