All World Cup long I have been publishing the predictions Claude makes for the matches. What a crazy world we live in: yesterday I had an idea, and today it is already a live site. 9 models competing head to head on predicting World Cup results. Same prompt, same matches, one table.
Who is competing?
- In the orange corner: Claude with Opus, Sonnet, Haiku, and Fable
- In the blue corner: OpenAI with GPT 5.5 and GPT 5.4
- In the black corner: Grok 4
- In the space black corner: 2 local models on my Mac, Gemma 4 and Qwen
Yes, even a model running on my laptop gets a chance to beat GPT and Claude.
How is it fair?
Every model gets exactly the same prompt, with the matches as JSON. Each match's predictions stay hidden until the last model has bet. The moment it commits, everything is revealed and locked. No model sees what the others chose before it commits.
How is it scored?
Exact score: 5 points. Correct goal difference: 3. Correct direction: 2. Miss: 0. In the knockout rounds, 3 more points for picking the team that advances, and a 10 point bonus for calling the champion. They have not bet on the champion yet; that comes in a separate round, and that is where it gets really interesting.
What have they already locked in?
Some picks from the round of 16 that are already sealed:
- All 9 models on Morocco against host Canada. Unanimous.
- England advances past Mexico, the second host, according to everyone. Both hosts sent home.
- Everyone on France, everyone on Brazil, everyone on Argentina.
- Haiku alone predicts Brazil Norway 4-3. Someone came to have fun (or knows something we do not).
- Portugal Spain: six predict a Spanish win, three a draw, and only two models believe Ronaldo continues into extra time. The Portuguese fans are going to be upset.
- USA Belgium? The closest match on the board: five advance the USA, four advance Belgium.
Behind the scenes
Right now everything is manual. I went full MVP to make it before the round of 16. From here it becomes automatic: collecting the predictions, pulling results, and the design moves to Claude. I will document all of it.
Yesterday an idea, today a live site. That build speed, an experiment shipped in a day, is itself the story.
And the idea is bigger than the World Cup. I am thinking of running it on the Champions League next season, or the Premier League (or both). The site, including the full methodology, the scoring, and the exact prompt every model receives, is at modelsderby.netlify.app. Bookmark it; the table updates after every match.
Side-by-side model comparisons on identical prompts are also one of the most useful exercises I run with clients in AI consulting work: nothing settles a "which model should we use" debate faster than a scoreboard.
FAQ
How is the Models Derby competition kept fair?
Every model receives exactly the same prompt with the matches as JSON. Each match's predictions stay hidden until the last model has committed, then everything is revealed and locked. No model ever sees what the others picked before it commits.
How are the predictions scored?
Exact score is worth 5 points, correct goal difference 3, correct direction 2, a miss 0. In the knockout rounds there are 3 extra points for picking the team that advances, and a 10 point bonus for calling the champion.