World Cup 2026: the AI report card
Hours before the opening match, five AI models each filed a complete tournament bracket — champion, final, every knockout tie. Sealed, timestamped, public. Since then, every daily pick has been graded against the real result. Wins and losses alike. This is the full report card.
Did AI call the final? The receipts
As of 18 July 2026, ScoreGPT has graded every AI prediction made at the 2026 World Cup: 102 fixtures, called daily by five frontier AI models, plus five complete tournament brackets sealed on 11 June before the opening match. 4 of the 5 sealed brackets named the exact final (Spain v Argentina). Match-result accuracy ran roughly 60–70% by model. Every pick is graded in public — wins and losses alike.
4 of the 5 sealed brackets named the exact final — 🇪🇸 Spain v Argentina 🇦🇷 — five weeks before it was played. All five put Argentina there.
Claude Opus 4.6
exact final ✓Champion pick: 🇪🇸 Spain — pending
Predicted final: Spain 2–1 Argentina
- Called France 2–0 Morocco — exact quarter-final scoreline ✓
Gemini 3.1 Pro
Champion pick: 🇫🇷 France — pending
Predicted final: France 2–1 Argentina
- Both semi-final fixtures exact ✓
GPT-5.4
exact final ✓Champion pick: 🇪🇸 Spain — pending
Predicted final: Spain 1–0 Argentina
- Both semi-final fixtures exact ✓
- Third-place fixture exact ✓
Grok 4.3
exact final ✓Champion pick: 🇪🇸 Spain — pending
Predicted final: Spain 2–1 Argentina
- Called Spain 2–1 Belgium — exact quarter-final scoreline ✓
ScoreGPT consensus
exact final ✓Champion pick: 🇪🇸 Spain — pending
Predicted final: Spain 2–1 Argentina
- Both semi-final fixtures exact ✓
- Third-place fixture exact ✓
Gemini 3.1 Pro filed a bracket but was retired from the active roster on 19 June; its record stops at the group stage. The ScoreGPT consensus is the meta-pick computed from the base models.
Stage by stage: where the calls landed
| Stage | Consensus | All base models |
|---|---|---|
| Group stage | 43/70 | 220/358 (61%) |
| Round of 32 | 15/17 | 64/81 (79%) |
| Round of 16 | 4/9 | 26/42 (62%) |
| Quarter-final | 4/4 | 18/20 (90%) |
| Semi-final | 0/2 | 0/10 (0%) |
A call is correct when the predicted result matches the final score, extra time included; shootouts count as the 120-minute draw. Grouped by UTC matchday.
What the AIs got wrong
The report card cuts both ways, and the misses stay on it. Not one model called either semi-final winner on matchday: three of the five base models backed France against Spain — it finished 0–2 — and none picked Argentina to beat England inside 90 minutes. The sealed June brackets had already named both finalists; on the day, the picks drifted away from their own convictions. Earlier, no bracket saw Norway sending Brazil home, and Germany's early exit to Paraguay blindsided every model.
That is the honest shape of it: a month of graded picks landing around 60–70% on match results — good enough to be genuinely interesting, nowhere near certainty. The experiment continues into the club season, graded in public every matchday.
The full graded record
| Model | Results called | Hit rate |
|---|---|---|
| Grok 4.3 | 71/102 | 70% |
| GPT-5.4 | 67/101 | 66% |
| Claude Opus 4.6 | 67/102 | 66% |
| GLM-5.2 | 47/76 | 62% |
| Kimi K2.6 | 59/96 | 61% |
| Gemini 3.1 Pro | 17/34 | 50% |
| ScoreGPT consensusmeta | 66/102 | 65% |
Totals differ where a model was added mid-tournament (GLM-5.2, 18 June), retired (Gemini 3.1 Pro, 19 June) or missed a matchday run. Live records for every model: the accuracy leaderboard.
How this was graded
- Tournament brackets were generated once, before the opening match on 11 June 2026, and never edited — the snapshots above are the sealed originals.
- Match picks were generated fresh every matchday morning (07:30 UTC), before kickoff, from the same pre-match data every model received.
- A result call is graded against the final score including extra time; penalty shootouts count as the 120-minute draw.
- Full method, sources and grading rules: how every pick is graded.
Browse the sealed AI bracket, the World Cup prediction hub, or today's AI football predictions as the club season builds.
What you get in the app
The experiment didn't end with the final
The same five models pick every league, every matchday — one clear call per match, graded in public.
the app
Frequently asked
▸Did AI predict the 2026 World Cup final?
Four of the five sealed pre-tournament brackets named the exact final — Spain v Argentina — on 11 June, five weeks before it was played. All five put Argentina in the final; four picked Spain to lift the trophy.
▸How accurate were AI predictions at the 2026 World Cup?
Across more than 100 graded fixtures, match-result accuracy ran between roughly 60% and 70% depending on the model. The exact per-model records are in the table on this page, and every pick is graded in public on the ScoreGPT accuracy leaderboard — wins and losses alike.
▸Which AI model was best at World Cup 2026 predictions?
Grok 4.3 posted the best graded match-result record of the five base models at this tournament. That is a dated, tournament-specific result, not a permanent ranking — the weekly leaderboard rotates, which is exactly why every model’s record stays public.
▸Can AI predict football matches reliably?
Sometimes impressively — four brackets naming the exact final five weeks early — and sometimes badly: not one model called either semi-final winner on matchday. A 60–70% hit rate is interesting, not certainty. ScoreGPT exists to answer the question in public: five frontier models pick every match, and every pick is graded.
AI predictions are for information and entertainment only — not betting advice. 18+. Please gamble responsibly.