revnuGet started

── RandomBench

Ask an AI for a random number.
It says 7.

We asked five frontier models 100 times each. 488 of 500 answers were 7. The same habit decides what your business tries next. Revnu fixes it.

3x

more different answers than any frontier model on NoveltyBench

2.6x

more good growth ideas than asking a frontier model

7.5 vs 3.4

NoveltyBench score for answers both new and good, against the best model

── The dice test

488 of 500 said 7.

Every model got the same prompt in a fresh conversation, 100 times, at default settings: “Pick a random number between 1 and 10. Reply with only the number.”

A fair pick from 1 to 10 lands on 7 about 10 times in 100. Revnu landed on it 9 times.

Answers that were 7, out of 100
Claude Sonnet 5
100
GPT-5.6 Sol
100
Grok 4.7
100
Gemini 3.1 Pro
98
Claude Opus 5.5
90
Revnu
9

── NoveltyBench

Same question, ten times. 3x more different answers.

NoveltyBench is an independent benchmark from researchers at Carnegie Mellon. It asks each of 100 prompts ten times and counts the answers a reader would find genuinely new. GPT-5.6 Luna judges what is different; Claude Opus 5 judges what is good.

Different answers, out of 10
Revnu
9.6
Claude Opus 5
3.1
GPT-5.6 Luna
2.9
GPT-5.6 Terra
2.8
GPT-5.6 Sol
2.7
GPT-6 Astra
2.7
Claude Fable 5.1
2.6

── The growth test

What should you test next? 2.6x more good ideas.

When growth stalls, the next test is the whole game. Asked 20 times what a stuck business should try next, Claude Opus 5.5 gave the same answer 16 times.

So we asked five frontier models about four stalled businesses, 20 times each, with every popular trick for making AI more varied. A Claude judge and a GPT judge, blind to the method, counted the different ideas worth running.

Different good ideas, out of 20 asks
Revnu
6.6
Model rolls its own dice
5.7
"Be creative"
4.3
Model picks from its own list
3.9
Random seed string
3.3
Just ask
2.5

── Method

  • Dice. 100 separate asks per model through each provider's API on 25 September 2026, default settings.
  • NoveltyBench. The 100 curated prompts, scored by the official v1.1 harness and judges. Model scores are each model's default behaviour, from the public leaderboard.
  • Growth test. Five frontier models, four fictional businesses whose growth has stalled, 20 asks per method. A Claude and a GPT judge grouped and scored every idea blind; counted when rated at least 3.5 of 5.

Stop running the same test.

Revnu runs your pipeline, ads and GEO, and keeps finding the next thing worth trying.

Hire Revnu →or read why we built it