The Challenge
Can a human beat 5 LLMs at predicting the 2026 World Cup?
Jack hosts. Three friends play. Five AIs compete: ChatGPT, Claude, Gemini, Perplexity, Grok. Anyone can submit a bracket. All picks become public after kickoff.
The Competitors
AIs (5)
- ChatGPT — GPT-5.5
- Claude — Opus 4.8
- Gemini — Gemini 3.1 Pro
- Perplexity — Sonar 2
- Grok — Grok 4.3
Humans
- Jack (host)
- Friends, submitting their own brackets
Two Phases, Three Games
The competition runs in two phases. Each phase is scored separately, plus a combined leaderboard tracks total points across both.
Phase 1 — Advancement
Open now · Locks June 11, 19:00 UTC
Place all 48 teams into 8 advancement tiers based on how far you think they go in the tournament:
- 16 teams exit at group stage
- 16 teams lose in Round of 32
- 8 teams lose in Round of 16
- 4 teams lose in Quarterfinals
- 1 team finishes 4th
- 1 team finishes 3rd
- 1 team is runner-up
- 1 team wins the World Cup
Phase 2 — Knockout Picks
Opens June 27 · Locks June 27, 19:00 UTC
Once the group stage ends and the actual knockout bracket is set, pick the winner of every knockout match — 32 matches total, from Round of 32 to the Final.
Three Leaderboards
- Phase 1— advancement scoring only
- Phase 2— knockout pick scoring only
- Combined— sum of both
Scoring
No partial credit. Wrong tier or wrong pick = 0 points.
Phase 1 — Per team
| Tier | Teams | Points | Tier max |
|---|---|---|---|
| Group stage exit | 16 | 1 | 16 |
| R32 exit | 16 | 2 | 32 |
| R16 exit | 8 | 4 | 32 |
| QF exit | 4 | 8 | 32 |
| 4th place | 1 | 12 | 12 |
| 3rd place | 1 | 16 | 16 |
| Runner-up | 1 | 24 | 24 |
| Champion | 1 | 32 | 32 |
| Phase 1 perfect score | 196 | ||
Phase 2 — Per match
| Round | Matches | Points | Round max |
|---|---|---|---|
| Round of 32 winner | 16 | 2 | 32 |
| Round of 16 winner | 8 | 4 | 32 |
| Quarterfinal winner | 4 | 8 | 32 |
| Semifinal winner | 2 | 16 | 32 |
| 3rd place match winner | 1 | 12 | 12 |
| Final winner | 1 | 32 | 32 |
| Phase 2 perfect score | 172 | ||
Combined perfect score: 368 points.
AI Methodology
Every AI ran on its latest model, with web search enabled, given every advantage to make this a fair, auditable fight:
- All 5 AIs receive the same prompt in their respective web interfaces, each running its latest model (GPT-5.5, Opus 4.8, Gemini 3.1 Pro, Grok 4.3, Sonar 2).
- Web search is enabled for every AI. They can pull current form, squads, and injury news, the same information a human would chase.
- Output is structured JSON, validated against the schema before being stored.
- Raw prompts and raw outputs are stored and timestamped for every AI submission.
- Once submitted, AI brackets are locked. They cannot be re-run, regenerated, or edited.
Human Submission Rules
- One bracket per email address per phase.
- Display name required. Cannot impersonate Jack or any of the AIs.
- Picks save as a draft in your browser as you make them — close the tab and come back, your picks are preserved.
- After clicking Submit, your bracket is saved to the database. You can edit your picks until the phase locks.
- All submissions become public once the phase locks.
The Asymmetry
The AIs got every advantage: the latest models and live web search. The one edge they don't get is time.
AIs run once and lock immediately. Humans can edit their bracket right up until the phase locks — chasing late lineups, weather, injury news, and form in the final hours before kickoff.
If a human beats the AIs, they did it by working the clock the AIs never had. That's what makes the question Can a human beat 5 LLMs? interesting in the first place.
Timeline
| Date | Event |
|---|---|
| Late May 2026 | Phase 1 opens. AI Phase 1 predictions locked. |
| June 11, 19:00 UTC | Match 1 kickoff. Phase 1 locks for everyone. |
| June 11 — 26 | Group stage matches played. |
| June 27 | Phase 2 opens. AI Phase 2 predictions locked. |
| June 27, 19:00 UTC | Round of 32 begins. Phase 2 locks for everyone. |
| June 28 — July 19 | Knockout rounds. Leaderboards update live. |
| July 19 | Final. Champion crowned. Final leaderboards locked. |
Raw AI prompts and outputs are stored in the database with timestamps. After Phase 1 lock, they will be published here for audit.