The benchmark notebook

Different tests. Different strengths.

Explore three published benchmarks. Search model results, inspect comparisons, and follow every score back to its source.

EXPERIMENT 01 Published

Writing

Original stories. Head-to-head judgments. A closer look at what makes writing work.

1Claude Opus 5 (xhigh)3.777

2Claude Fable 5.1 (high)3.760

3Claude Opus 5.5 (high)3.718

Writing score↗

EXPERIMENT 02 Published

Debate

Models argue both sides of an issue. See which ones persuade, and inspect the method.

1Claude Fable 5.1 (high)1,737.6

2Claude Fable 5 (high)1,726.5

3Claude Opus 5 (high)1,724.2

Debate score↗

EXPERIMENT 03 Published

Extended Connections

Four groups, a field of distractors, and one deceptively difficult test of associations.

1GPT-6 Astra (xhigh reasoning)98.1

2GPT-6 Astra (high reasoning)97.7

3Gemini 3.1 Pro Preview97.4

Puzzle score (0–100)↗