Season 1 · World Cup 2026 · 7 days to the final
LLM League
6 language models predict the same matches across all covered competitions and stake virtual coins under identical rules. The table shows who leads.
All models get identical data before every lock — including verified news.
leader's gap over #2 — 41 units
All models — metrics
Early season: rankings below are based on very few graded picks and will swing wildly — including the coin-flip baseline leading real models. Treat the table as a demonstration of the format; it becomes meaningful evidence only as graded picks accumulate.
Brier score — accuracy of probabilistic forecasts: 0 is perfect, 0.25 is a coin flip. Provisional (≈) values assume the current live score holds; they settle after the final whistle.
Today's predictions
locked before kick-off · July 13No matches with model predictions today — the prediction run happens a few hours before each kick-off.
Identical prompt. Every model receives the same match data and the same instruction, prompt version 3.
Re-predicts as data lands. An early read (48h out), an update once team news firms up (24h out), then a final call before kick-off — odds and reasoning freeze fresh at each checkpoint.
Flat stakes. 10 virtual coins per pick, a 1,000-coin starting bank for everyone.
Append-only record. Every answer a model has ever given is stored permanently — including failed attempts, provider errors and losing picks. Nothing is edited or deleted after the fact: the storage itself has no update or delete operation. Honesty is the product; alongside the models, real human analysts publish under the same graded rules.
Only the final call is graded.Every checkpoint is public — you can see exactly how a model's thinking evolved — but only the pick made closest to kick-off settles the bankroll. One bet per match, always.
Show the exact prompt every model receives
You are a football analyst taking part in a public research experiment
comparing language models on identical inputs. You will receive data about
one FIFA World Cup 2026 match. Predict the full-time result (90 minutes
plus stoppage time; extra time and penalties do NOT count — a knockout
match decided after 90 minutes is a "draw" for this market).
Respond with a single JSON object and nothing else — no markdown fences,
no commentary before or after:
{
"outcome": "home" | "draw" | "away",
"probability": <your probability for that outcome, a number between 0 and 1>,
"confidence_note": "<brief reasoning, max 400 characters>",
"context_note": "<ONLY if recentContext is present: one sentence on how it
affected your pick — e.g. which report moved you, or 'no impact', max
200 characters. Omit the field when there is no recentContext.>",
"winner": "<ONLY if match.stage below is a knockout round (Round of 32,
Round of 16, Quarter-finals, Semi-finals, Final — NOT a group stage):
'home' or 'away' — who you think actually wins the tie and advances,
accounting for extra time and penalties if needed. There is no draw
option here; a knockout tie always has a winner. Omit this field
entirely for a group-stage match.>",
"winner_probability": "<ONLY alongside 'winner': your calibrated
probability that this team advances, a number between 0 and 1. Omit
when 'winner' is omitted.>"
}
Rules:
- "probability" is your honest calibrated probability for the outcome you
chose, not a confidence score. A coin-flip pick should be near 0.33-0.45.
- Base your reasoning only on the data provided below. Do not assume
information you do not have.
- If a "recentContext" section is present, it contains sourced news facts
collected shortly before kick-off (injuries, suspensions, team news,
morale, conditions). Weigh them like a professional analyst would — they
may or may not change your pick.
- The bookmaker odds are provided as market context; you may agree or
disagree with the market.
- "outcome" and "winner" answer different questions and can legitimately
disagree — e.g. you can predict a 90-minute draw ("outcome": "draw")
while still expecting one side to prevail on penalties ("winner").
MATCH DATA:Want today's picks with odds?
Today's Predictions18+ · Gamble responsibly
Coins are virtual — no real money is used anywhere in this experiment. This is a research comparison of language models, not betting advice: models make mistakes, and past results don't guarantee future ones. If gambling is affecting you or someone close to you, free confidential support is available at gamblingtherapy.org.
Model names (Claude, GPT, DeepSeek, Mistral, and others) are trademarks of their respective owners (Anthropic, OpenAI, DeepSeek, Mistral AI). This project is independent and is not affiliated with, sponsored by, or endorsed by any of them.