The 2026 World Cup kicked off today — 48 teams, 104 matches, the biggest tournament in history.
I just switched my coding agent to a freshly released model called fable-5. Benchmarks are boring. So I gave it the most falsifiable task I could think of: predict the World Cup winner. Knockouts start in a month. The scoreboard doesn't negotiate.
What I fed it
No vibes — real opening-day data:
- Bookmaker odds (consensus backed by billions): Spain +450, France +500, England +700, Brazil +800, Argentina & Portugal +900
- Goldman Sachs' quant model: Spain 26%, France 19%, Argentina 14%
- One counterintuitive stat: in 15 World Cups from 1966 to 2022, the pre-tournament favorite lifted the trophy only 3 times
The reasoning chain mattered more than the answer
fable-5 didn't blurt out a name. It structured the problem in three layers:
1. Odds are the strongest prior. Betting lines aren't opinions — they're money voting. Spain at +450 implies roughly an 18-20% title probability after the vig. Any prediction that overrides the market needs information the market hasn't priced in.
2. The "favorite's curse" is math, not mysticism. Single-elimination football has enormous variance — penalty shootouts are near coin-flips. A 26% favorite is, equivalently, a team that fails to win 74% of the time. Three out of fifteen is exactly what that structure predicts.
3. This edition favors squad depth. The 48-team format adds an extra knockout round. North American summer heat and long flights turn fitness management into a hidden battlefield. Both shift weight from star ceiling to roster depth.
The verdict
Not a name — a distribution:
| Team | fable-5's probability |
|---|---|
| Spain | 22% (deliberately shaded below Goldman's 26% — "consensus itself is a risk") |
| France | 18% (best squad depth in the field) |
| Argentina | 13% (champion's mentality, aging core) |
| England | 11% |
| Brazil | 10% |
| Portugal | 8% |
| The other 42 | ~18% — highest dark-horse odds in World Cup history |
One line summarizes its stance:
"If you must pick one name, pick Spain. But if you understand probability, the real answer is: no team is more likely to win than not to win."
Why this stuck with me
Ask most people who'll win and you get a name plus emotions. Ask a good model and you get a prior, adjustments, confidence levels — and an honest "I'm probably wrong, 78% of the time."
That's the thing that's actually scarce in the AI era. Not answers — the structure of judgment. Anyone can output an answer. Putting uncertainty on the table, unprompted, is what separates a tool you can trust from a tool that flatters you.
I'll revisit this piece in a month and grade it publicly. If Spain wins, credit the model. If not — well, it told you so: even Spain loses this tournament 74% of the time.
Data: BetMGM / DraftKings / FanDuel opening-day lines; Goldman Sachs quant model; ESPN.
Top comments (0)