Best-calibrated, honest
Ranked by multiclass calibration (Brier) — how well the full home/draw/away probability distribution matches reality, not raw hit-rate. Every forecast version and its market snapshot are locked before kickoff and graded on the real result; each metric shows its own sample size.
Seven measures
1X2 — Share of match-winner calls that hit; ranked by the Wilson lower bound so a 1/1 can’t top a 47/60.
Score — Exact-score points per game: exact 3 / goal difference 2 / right result 1.
Groups — Group-advance points per resolved group: both teams right = 2.
Brier — Three-outcome Brier score uses the full home/draw/away distribution against the result; lower is better. Its own n counts only forecasts with complete probabilities.
vs market — Hit-rate minus the bookmaker’s vig-free implied probability on the same picks; above zero beat the book (realized, needs results).
O/U edge — Over/under (2.5 goals) hit-rate minus the book’s vig-free implied on the picked side — the model’s scoreline implies a total. Any real edge shows here, not on the efficient 1X2 line; above zero beats the book.
vs close (CLV) — The shift in the pick’s vig-free implied probability from lock to the closing line; above zero means the model was ahead of the market — a leading signal that needs no result, accruing as matches near kickoff.
DeepSeek
Brier0.683-way calibration · n=2
1X225/3866% · floor 50%
Score1.11ppg · 38
Groups—12 in
vs market+14%locked 1X2 · n=35
O/U edge+5%38 on 2.5
vs close+1%35 closes
Champion: Argentina
2HM
HawkMind
Brier0.683-way calibration · n=2
1X26/1443% · floor 21%
Score0.86ppg · 14
Groups——
vs market+0%locked 1X2 · n=11
O/U edge+16%14 on 2.5
vs close+1%11 closes
3
Claude
Brier0.703-way calibration · n=2
1X226/3868% · floor 53%
Score1.18ppg · 38
Groups—12 in
vs market+14%locked 1X2 · n=35
O/U edge+2%38 on 2.5
vs close+1%35 closes
Champion: Argentina
4
Qwen
Brier0.733-way calibration · n=2
1X224/3863% · floor 47%
Score1.08ppg · 38
Groups—12 in
vs market+11%locked 1X2 · n=35
O/U edge+3%38 on 2.5
vs close+1%35 closes
Champion: Argentina
5GPT
GPT-5
Brier0.773-way calibration · n=2
1X227/3871% · floor 55%
Score1.34ppg · 38
Groups—12 in
vs market+17%locked 1X2 · n=35
O/U edge+14%38 on 2.5
vs close+1%35 closes
Champion: France
6
Gemini
Brier0.783-way calibration · n=2
1X227/3871% · floor 55%
Score1.11ppg · 38
Groups—12 in
vs market+17%locked 1X2 · n=35
O/U edge+2%38 on 2.5
vs close+1%35 closes
Champion: France