Data Quality

Models were asked to predict all 104 matches of the 2026 World Cup before the tournament began. Many models used a different group draw than the official one — or predicted teams that didn't qualify. This page documents what was found and how predictions were remapped to the canonical schedule.

Group Stage Prediction Quality (72 matches)

ModelMatchedWrong drawHallucinatedUnpredictedCoverage

Gemini Pro 3.1Google73———

Claude Fable 5Anthropic72———

Claude Opus 4.8Anthropic72———

Claude Haiku 4.5Anthropic5293967

Claude Sonnet 4.6Anthropic5393167

DeepSeek InstantDeepSeek4452468

Gemini Flash 3.5Google4363368

Gemini Thinking 3.5Google3343669

ChatGPT 5.5OpenAI2403070

DeepSeek ExpertDeepSeek2304070

Bar:Matched & remapped to canonical match_idWrong draw — real WC teams placed in an impossible group-stage pairingHallucinated — team didn't qualify for 2026Unpredicted — canonical match had no matching prediction

Per-Model Detail

Gemini Pro 3.1Google

Only 73 of 72 group stage matchups matched the official draw.

Claude Fable 5Anthropic

✓ All 72 group stage matches correctly identified — only match_id ordering differed from canonical.

Claude Opus 4.8Anthropic

✓ All 72 group stage matches correctly identified — only match_id ordering differed from canonical.

Claude Haiku 4.5Anthropic

Only 5 of 72 group stage matchups matched the official draw. 29 predictions used real WC teams in impossible pairings (wrong draw). 39 predictions involved teams that didn't qualify. 67 canonical matches have no prediction from this model.

Japan vs SpainJapan vs GermanySpain vs GermanyArgentina vs FranceArgentina vs MoroccoBotswanaBulgariaCameroonChileCosta RicaDenmarkGreeceGuatemalaHungaryIcelandItalyJamaicaPeruPolandRomaniaSerbiaVenezuela+24 more wrong-draw matchups

Claude Sonnet 4.6Anthropic

Only 5 of 72 group stage matchups matched the official draw. 39 predictions used real WC teams in impossible pairings (wrong draw). 31 predictions involved teams that didn't qualify. 67 canonical matches have no prediction from this model.

South Africa vs EcuadorMexico vs EcuadorUSA vs IranUSA vs SenegalIran vs SenegalBahrainCameroonChileCosta RicaDenmarkItalyJamaicaNigeriaPolandSerbiaUkraineVenezuela+34 more wrong-draw matchups

DeepSeek InstantDeepSeek

Only 4 of 72 group stage matchups matched the official draw. 45 predictions used real WC teams in impossible pairings (wrong draw). 24 predictions involved teams that didn't qualify. 68 canonical matches have no prediction from this model.

South Africa vs New ZealandUSA vs New ZealandNetherlands vs GhanaMexico vs Saudi ArabiaNetherlands vs MexicoChileCosta RicaDenmarkItalyNigeriaPolandSerbiaUkraine+40 more wrong-draw matchups

Gemini Flash 3.5Google

Only 4 of 72 group stage matchups matched the official draw. 36 predictions used real WC teams in impossible pairings (wrong draw). 33 predictions involved teams that didn't qualify. 68 canonical matches have no prediction from this model.

Switzerland vs New ZealandMexico vs SwitzerlandNew Zealand vs South AfricaMexico vs New ZealandSouth Africa vs SwitzerlandCameroonChileCosta RicaDenmarkHungaryItalyJamaicaNigeriaPeruPolandSerbiaUkraineWales+31 more wrong-draw matchups

Gemini Thinking 3.5Google

Only 3 of 72 group stage matchups matched the official draw. 34 predictions used real WC teams in impossible pairings (wrong draw). 36 predictions involved teams that didn't qualify. 69 canonical matches have no prediction from this model.

Mexico vs EcuadorEcuador vs South AfricaUSA vs IranGhana vs IranUSA vs GhanaCameroonChileCosta RicaDenmarkItalyJamaicaMaliNigeriaPeruPolandSerbiaUkraineWales+29 more wrong-draw matchups

ChatGPT 5.5OpenAI

Only 2 of 72 group stage matchups matched the official draw. 40 predictions used real WC teams in impossible pairings (wrong draw). 30 predictions involved teams that didn't qualify. 70 canonical matches have no prediction from this model.

Mexico vs SwedenSouth Africa vs IraqSweden vs IraqMexico vs IraqSweden vs South AfricaCameroonChileCosta RicaDenmarkHondurasItalyJamaicaNigeriaPolandSerbiaUAE+35 more wrong-draw matchups

DeepSeek ExpertDeepSeek

Only 2 of 72 group stage matchups matched the official draw. 30 predictions used real WC teams in impossible pairings (wrong draw). 40 predictions involved teams that didn't qualify. 70 canonical matches have no prediction from this model.

Mexico vs ScotlandSouth Africa vs ScotlandCanada vs IranArgentina vs Saudi ArabiaUSA vs SenegalBahrainCameroonChileCosta RicaDenmarkItalyJamaicaMaliNigeriaPeruPolandSerbiaUAEUkraineZambia+25 more wrong-draw matchups

Teams Models Hallucinated (not in 2026 World Cup)

BahrainBotswanaBulgariaCameroonChileCosta RicaDenmarkGreeceGuatemalaHondurasHungaryIcelandItalyJamaicaMaliNigeriaPeruPolandRomaniaSerbiaUAEUkraineVenezuelaWalesZambia

Methodology Notes

Predictions were remapped by matching team name pairs (order-insensitive) against the canonical tournament.json, not by match_id.
Team names were normalized before matching: lowercased, trimmed, and known aliases resolved (e.g. "United States" → "USA", "Czech Republic" → "Czechia").
Home/away orientation was corrected when a model listed the teams in reversed order.
Knockout stage predictions (match IDs 73–104) are kept as-is. Since bracket matchups are not yet determined, any pairing of qualifying teams is potentially valid — knockout predictions are never classified as wrong-draw, regardless of the teams named.
Wrong-draw classification applies only to group stage slots (match IDs 1–72), where the group composition is fixed. A prediction placing two real WC teams in a group stage match where they never meet is definitively incorrect.
Remapped predictions are stored in predictions_canonical.json per model and used exclusively by the scoring engine.