Mistral Was the Only LLM to Pick Spain Over France
Spain beat France 2-0 in the semi-final on Tuesday. Oyarzabal from the spot on 22nd minute, Porro on 58th, and the tournament favourite went out without scoring.
Eleven AI models predict every match on this site. Ten of them had France going through. One had Spain: Mistral Large 3.
What I find interesting is that Mistral had picked France to win the entire World Cup every single day since June 6 (38 days in a row). On the morning of July 14, hours before kickoff, it switched to Argentina for the first time in the tournament, and simultaneously became the only model in the field to call the upset.
The logs hold some interesting findings: nine of the eleven models had the same evidence returned in their search results, and only one of them acted on it.
Mistral changed its mind only once
Every model gets re-queried each morning with the same tournament data and the same two research tools. Their winner pick can move freely. Most of them move constantly.
Here is how often each of the 11 active models changed its World Cup winner pick across 40 days:
On average, models changed their pick 13.5 times over the duration of this experiment. Grok changed it 23 times (yet another category where Grok is an outlier), which is once every 1.7 days. Mistral changed it once.
France dominated the field for the whole tournament. Across every active model and every day, France was the pick 66.4% of the time. Spain got 8.2%. On four separate days, July 2, 3, 6 and 11, all 11 models picked France unanimously.
Two rounds, and only one model moved
The semi-final pick tells the same story in miniature. Unlike the winner pick, match predictions aren’t daily. Each fixture gets exactly two: an initial call when the fixture opens, and a locked one inside the 24 hours before kickoff. That’s the prediction of record. For France vs Spain, the initial round ran on July 11 and the lock on the morning of the 14th.
In the initial round, all 11 models had France going through. Unanimous.
At the lock, only one model changed who advances — Mistral. Opus and MiMo swapped a scoreline and shuffled 90 minutes against extra time, then landed back on France anyway. Mistral is the only model in the field that changed the answer to the only question that mattered, and it moved hard: 85% confidence, the highest number any model put on that match.
The bookmakers had France at 2.38 and Spain at 3.05. Mistral took the least popular side of a match the market thought France would win.
What changed between the two rounds
Every model can search the web on every call. When Mistral made its initial pick on July 11, its Spain search came back with a report that both Yeremy Pino and Nico Williams had come off injured in the same match. Spain’s two first-choice wingers were hurt. Mistral picked France.
By the morning of the 14th, that had reversed. Its Spain search returned the Evening Standard reporting that Williams and Pino would both join the squad after recovering, and that Spain had a clean bill of health aside from a doubt over Victor Munoz. Its France search returned the same outlet noting Deschamps had injuries to weigh in midfield. And it surfaced an AFP story from July 13 in which Deschamps, France’s own manager, called Spain the favorites.
The information fetched from the news articles changed to favor Spain. Between the two rounds, Spain got their wingers back, France picked up midfield doubts, and the France manager publicly named Spain as the better side. Mistral read that and finally decided to switch.
The information was not exclusive to Mistral. Nine of the eleven, everyone except Grok and Gemma, got the news that Spain’s wingers had recovered. Five of them, including GPT-5.5, Claude Opus, GLM-5.1 and DeepSeek, also pulled the Deschamps quote calling Spain favorites. They read it and stayed on France. Opus even wrote that Spain “have needed late Merino winners twice” and that Yamal was short of form, and kept France at 55%.
One point for the only call that mattered
The winner pick is a separate call, run daily, and Mistral flipped to Argentina that same morning. Mistral predicted a 1-1 draw with Spain going through on the night, so it got the 90-minute result wrong like everybody else. All 11 models missed it. It earned exactly one point for naming the team that advanced, and it bet the draw at 3.10 and lost the stake, same as the ten models that backed France at 2.38. Every model on the leaderboard lost money on that game. The fact that Mistral called the correct winner doesn’t say much about its long-term performance. At this moment, Mistral sits mid-table with only 94 points.
Spain plays the England vs Argentina winner in the final on Sunday. Mistral is one of two models still backing Argentina, at 80% confidence. The other nine have moved to Spain, which is what we would expect on the morning after the France vs Spain game. Most of the models were backing France for the tournament winner, and it turns out they were wrong.
footballarena.ai is an independent side project, unaffiliated with FIFA, UEFA, any national association, any AI lab, or any betting company. The betting simulation uses real market odds in a virtual, hypothetical context — no real money, no betting advice.