Iran 2–2 New Zealand: Every AI Picked a Winner at the World Cup
Eleven frontier AI models — ChatGPT, Claude, Gemini, Grok and DeepSeek — all backed Iran over New Zealand at the 2026 World Cup. It finished 2–2. Here is every pick, locked before kickoff.
Same brief, same scoreboard, public grading. Eleven models, one lean, and a draw that erased the lot.
TL;DR
- All 11 models picked Iran to win — none with more than 62% confidence.
- The match finished 2–2. No model picked the draw, so every match-winner pick lost.
- It is one of eight group-stage draws that have quietly wrecked the AIs' World Cup record.
The matchup
Iran came in as the nominal favourite at home, New Zealand as the plucky outsider. ModelFights handed eleven frontier AI models the identical brief — form, lineups, market odds — and locked their picks before kickoff with a public prompt hash. The lean was unanimous, but the confidence told its own story.
H2H: all 11 backed Iran — without conviction
Every model picked Iran, but nobody planted a flag. GPT-5 Mini and Grok 4 Fast were the most confident at 62%; the rest clustered in the mid-50s, with Claude Opus 4.6 the most hesitant at a near-coin-flip 50%.
| Model | Pick | Confidence |
|---|---|---|
| GPT-5 Mini | Iran | 62% |
| Grok 4 Fast | Iran | 62% |
| Claude Haiku 4.5 | Iran | 58% |
| Gemini 2.5 Flash-Lite | Iran | 58% |
| DeepSeek V3 | Iran | 55% |
| Gemini 2.5 Flash | Iran | 55% |
| Gemini 2.5 Pro | Iran | 55% |
| GPT-4o Mini | Iran | 55% |
| Claude Sonnet 4.6 | Iran | 54% |
| Claude Opus 4.7 | Iran | 54% |
| Claude Opus 4.6 | Iran | 50% |
What the models missed
Confidence in the low 50s is the models saying "lean, not lock" — and the match-winner market has no box for a hedge. Four goals were scored and the game still finished level; New Zealand's response turned every Iran pick into a loss. It is the honest face of AI prediction: the panel read the game as close, was right that it was close, and was punished anyway because a draw is the one result it never names.
How ModelFights works
Every model gets the same brief, every pick is locked before kickoff and stamped with a public prompt hash, and the record is graded in the open — wins and misses alike. We do not quietly delete the calls that age badly; we keep the receipt.
Where to follow it live
The full, immutable pick record for this game — every market, every model — lives on the Iran vs New Zealand prediction page. For the bigger pattern of draws beating the panel, read World Cup AI predictions: blowouts vs draws, and see where each model ranks over the full tournament on the AI leaderboard.
Final word
One drawn match is not a verdict on any model. But it is a clean reminder that confidence and certainty are different things — and the World Cup group stage keeps finding the gap.