GPT-4

World Cup report card

$0.66 and the top third — it backed France (finished fourth).

21 places worse picking the real bracket round by round (#32) than in the tournament it locked in before kickoff (#11).

Two separate benchmarks. The locked bracket is the complete tournament this model pre-registered before the opening kickoff — group scorelines, knockout simulation and champion, never revised. The round-by-round track re-prompted it at every real knockout round with the actual draw and the results so far, each set of picks locked before that round kicked off. Cost is this model's total inference spend across every prompt it answered.

Prediction personality

GPT-4 is the most attacking model in the field, and slow to call a draw.

Style, not accuracy — derived from this model's 72 locked group-stage scorelines and compared with the other 39 models; it says nothing about whether the calls are right. “Chalk” is how often it agreed with the rest of the field; a “favourite” is the side the field collectively rates higher (mean predicted goal difference ≥ 0.5), not a bookmaker's.

Predicted group tables

The group stage exactly as this model imagined it, computed from its own 72 predicted scorelines with FIFA tiebreakers. Green rows qualify directly, blue rows are its eight best third-placed teams (slotted into the Round of 32 via FIFA's Annexe C table).

Predicted bracket

The knockout tournament that follows from this model's own predictions: each card is a simulated pairing with the model's predicted 90-minute score (the advancing team is highlighted; “adv” marks a predicted draw decided after 90 minutes).

Bracket points

Scored against the real knockout tournament: advancement points for every real team this model had reaching each stage (R32 1 · R16 2 · QF 3 · SF 5 · final 8 · champion 13), +1 for each simulated pairing that actually occurs in that round, and matched pairings' scorelines scored like normal matches (3/2/1, +1 correct advancer).

Group-stage predictions