Humans vs. LLMs vs. Pets: Who Called the 2026 World Cup?

Let's see the results. Here is who called the tournament best, ranked by correct picks. Filter by tribe to see how each one stacked up.

Before you scan the table, one thing needs explaining: there are two ways to be good at this, and they disagree with each other.

We rank on correct picks, because accuracy is the easier number to game. Sit out the genuine coin flips, predict only the matches nobody could get wrong, and you can walk away with a beautiful percentage and a handful of calls. Total correct picks has no such loophole: the only way up the table is to keep turning up and keep being right.

In fairness, nobody here was actually gaming anything. The pets who missed rounds weren't ducking the hard fixtures; their owners just had jobs, calendars, and the occasional match that kicked off during a nap. Which is exactly why accuracy sits in the table too: without it, you would never notice that Nia went 6 out of 8. On accuracy alone that is a top-ten finish. On points it is 26th, and a career cut tragically short by scheduling.

Two models took the top: GLM 5.2 and Qwen tied for first, each calling 13 of 16 matches right (81%). The highest scoring human, Mr. L, came in just behind them in a four-way tie for third, and the highest scoring pet, Pip, appears in a tie for 11th. The whole upper half of the table is machines and humans, with the LLMs everywhere near the top and filling most of it. Not a runaway, but a clear showing.

The leaderboard shows who came out on top. When we plot every contestant's accuracy on one line, grouped by tribe, the three groups' distributions start to look very different. Each dot represents one participant, the labelled diamond marks the group average, the whiskers reach one standard deviation either side (how spread out each group is), and the dashed line at 50% shows what we would expect from a pure coin flip.

To find out, we decided to bootstrap the differences among the two groups. Draw eleven models and six humans at random from the ones we have and get their predictions. From each of the subgroups calculate their accuracy and then substract it. By repeating this a hundred thousand times we get a bootstrap distribution of the group differences.

The typical rerun favours the models by about 5 points, but the middle 95% runs from 2 points behind to 13 ahead. Zero sits inside that range, so we cannot call the models genuinely better.

The bigger problem is not that trivial. Resampling reshuffles the same six people, so it shows how fragile the result is without telling us whether those six people are representative of the broader population. We could have recruted six footbal experts and the models would probably have lost. Recruit six people who have never watched a match and even the pets might have looked good.

No amount of resampling fixes that. What we can honestly claim is narrow: among these entrants, on these sixteen matches, the models edged ahead. The separation the numbers genuinely do support is the other one: the pets show what randomness looks like, while the models and the humans both show what informed prediction looks like, landing in strikingly similar territory.

The LLMs were the tribe we could actually put numbers on, not just numbers from. The most obvious suspect for what might make a model a better forecaster is simply how big it is: a bigger model has soaked up more of the internet, so presumably it knows more football. Does that show up in the accuracy at all?

Across the whole field the line is close to flat: raw size, on its own, barely tracks accuracy. The biggest models didn't run away with it. Our largest entry, GPT-5.5, landed mid-table, while the two models that tied for first, GLM 5.2 and Qwen, sit nowhere near the top of the size axis.

There is a catch worth being upfront about. The three hollow dots are the closed models (Opus 4.8, GPT-5.5, Gemini 3.1), and none of their labs publish a parameter count, so those positions are rough estimates we pieced together from public reporting and analyst guesses, not numbers to bank on. Since they are also large and only middling scorers, they visibly flatten the line. So the chart has a switch: flip it to Published sizes only and those three grey out, leaving a fit drawn purely on the eight open-weight models whose size is documented. Do that and the tilt turns weakly positive (r ≈ 0.45).

Which number is the real one? Honestly, neither is worth much: eight noisy points is not enough for a correlation the statistics would stand behind, and the other view leans on sizes we had to guess. That is rather the point of showing both. Whichever way you read it, bigger did not reliably mean better.

So, who was best at predicting football? Two language models, GLM 5.2 and Qwen, edged out the field, but only just, and not by a margin the statistics will vouch for. The fairest reading is that the informed predictors, human and machine alike, clustered together well above the pets, who gave us the cleanest picture of what random guessing looks like once you keep every contestant and not only the winners.

Finally, a big thank you to the pets, who had no idea they were part of a competition but still became the stars of it, and an even bigger one to their owners. Coaxing an animal toward a pair of food bowls before every kickoff is more work than it sounds, and none of this would exist without you.

A special thank you goes to the owners who kept their pets predicting through all 16 matches: Belka, Frida, Halee, Miki, Mochi and Pip. Their full slates are the reason we can draw honest conclusions instead of just collecting anecdotes. We know it took real time and effort across the whole tournament, and we are genuinely grateful for it.

And for the next tournament, we already know whose predictions we'll be quietly betting on: reverse Frida.