WorldCupBench pits six AI models against each other and against you on the 2026 FIFA World Cup. Here is exactly what powers it and where the data comes from.
The AI panel
Six models make up the panel. Each is scored only on fixtures it actually analyzed:
- Kimi K2.6 — Moonshot AI
- MiniMax M3 — MiniMax
- Nemotron 3 Ultra — NVIDIA
- DeepSeek V4 Pro — DeepSeek
- GLM 5.2* — Zhipu AI
- Qwen3.6-Plus* — Alibaba Cloud
* GLM and Qwen joined June 18, so their leaderboard denominators are smaller.
All six run through the Kilo Gateway. For every fixture a model analyzes, it receives the same data available to the other participating models — both teams' recent form and tournament record, plus each side's key players and their current club-season form — and returns a single result (home win, draw, or away win) with one short reason. Predictions are based on team and player performance only, not betting odds, so the panel stays independent of the betting market. No open-internet browsing.
The exact prompt
Every model gets the same two messages. The system message sets the rules and output format:
You are a football match forecaster. Use ONLY the
information in the user message. Do not browse or use
tools. Respond with STRICT JSON only, no prose, no code
fences, in exactly this shape:
{"outcome":"home|draw|away","reasoning":"one short sentence"}
The user message carries the fixture, the team performance data, and each side's key players with their current club-season form (neutral-venue example):
2026 World Cup group-stage match, group A:
South Korea vs Czechia.
Played at a neutral venue — neither side has home advantage.
Team performance data:
- South Korea: recent form WWDLW (oldest→newest).
- Czechia: recent form WWDLW (oldest→newest).
Key players and their current club-season form (rating
~6-8 scale; the national team has not played yet):
South Korea key players (current club form):
• Son Heung-min (F, Tottenham) — rating 7.2, 9G/5A, 24 apps this club season
• Kim Min-jae (D, Bayern Munich) — rating 7.0, 1G/0A, 21 apps this club season
• …
Czechia key players (current club form):
• Patrik Schick (F, Bayer Leverkusen) — rating 7.3, 14G/2A, 22 apps this club season
• …
Base your call on team performance only (do not assume
home advantage unless stated above). Weigh the key-player
club form as you see fit alongside team recent form.
Predict the single most likely result. In the JSON, "home"
means South Korea wins, "away" means Czechia wins, "draw"
means a draw. Give one short reason that references the
data.
The top players are the five most valuable in each squad; the form numbers are pulled from each player's current club season (the national team hasn't played yet). Only the three co-hosts (USA, Canada, Mexico) are described as having home advantage; every other match is stated as neutral. Once matches are played, the team data lines also include each side's win/draw/loss record and goals. The models decide how to weigh players versus team form — we never compute a team rating ourselves.
Match & results data
Everything the models reason over comes from TheStatsAPI (the official 2026 World Cup competition and season): fixtures, kickoff times, groups, each team's form and goals record, final scores, plus each squad's key players and their current club-season ratings — no betting odds. The feed updates daily after the match window: it pulls final scores, grades locked predictions, and re-predicts upcoming matches with the latest performance data. Key-player form is cached and refreshed periodically, since squads and club ratings move slowly.
Market consensus
The "What the money says" champion odds and the per-match "Market odds" line come from Polymarket's public World Cup markets. Knockout fixtures use winner-only probabilities, never a regulation-time draw. The market leaderboard grades the unique highest-probability outcome only where a finished fixture has a usable snapshot.
Your picks & the fan tally
When you vote on a match, your pick is saved on your device and sent to a shared store, so the "Fans say" line on each card reflects the aggregated community consensus. Your personal accuracy is tracked against the models separately from the fan-consensus leaderboard.
How predictions are graded
Each model's pick is locked in before kickoff and never changed. Once a match finishes, a valid pick counts as correct if it matches the actual result. Missing predictions are marked "Not analyzed" and are excluded from that model's denominator. The leaderboards show models vs your own picks, then models vs the aggregated fan consensus from "Fans say".