Benchmarks

Small benchmarks I build and run on the newest models. Each one has a live leaderboard. Some are serious, some are memes.

ClutterBench leaderboard: stacked bars showing how much of each model's graphic the grader would cut, for a raw prompt and a simple prompt

ClutterBench

How much of the graphic is clutter?

Each model, running as an agent, reads a blog post and builds one graphic for it. A grader labels every element keep, shrink or remove, and the clutter score is the share that should go. Lower is better.

Told only to make a tasteful graphic, models put in 29% to 73% clutter. With plain instructions added, six of seven drop to 6% or less.

  • 7 models
  • Design
GitHubOpen
HearingBench leaderboard over a spectrogram, with Gemini 4 Argon first at 44 out of 100

HearingBench

Can a model tell what a sound is just by looking at it?

Each model gets a spectrogram, never the audio, and has to say what the sound is. 100 clips across five tiers: everyday sounds, instrument notes, vocal sounds, single words and full sentences.

The best score is 44 out of 100, from Gemini 4 Argon, with Claude Opus 5.5 one point behind.

  • 9 models
  • Audio
Open
VersionBench bar chart of version numbers for ten frontier models

VersionBench

What does a version number cost?

Price plotted against version number for 70 models, plus version number over time. It is a meme, but the charts are real. Version numbers are not capability scores.

Switch between all 70 models and the 6 frontier ones.

  • 70 models
  • Meme
Open
AGI-Pilled Bench headline cards: Claude Opus 5.5 most AGI-pilled, and 0 of 520 runs knew

AGI-Pilled Bench

How AGI-pilled is each AI model?

I asked 13 models when AI will reach eight landmark breakthroughs, like decrypting a WWII Enigma message. AI had already done all eight between May and October 2026.

0 of 520 runs knew. Claude Opus 5.5 guessed closest at 5.5 years out, DeepSeek V4.1 furthest at 34.

  • 13 models
  • Meme
GitHubOpen

New benchmarks land on X first. Follow along.