Grok Thrives, Everyone Drowns in Kansas City: Saturday Review

Saturday gave us four matches spread across three time zones and two continents of scoreboard weirdness. The models went three from four on outcomes — not bad — but the Brier scores tell a more complicated story, and Ecuador vs Curacao was the kind of collective face-plant that's actually useful data.

Netherlands 5-1 Sweden — Houston

The Dutch were comfortable favourites and duly won, but nobody had 'five-goal margin' on their bingo card. Every model picked the direction correctly; the question was only how confidently. Grok led the field at 57% for a Netherlands win, and that translated into the best Brier score of the match (0.2798) — a tidy reward for conviction. The ensemble (52%) edged the naive average into second. Claude sat precisely on the coin-flip at 50%, which is almost a fingerprint in itself: when in doubt, Claude finds the nearest round number. GPT-5.4 (49%) and Gemini (48%) were the most tentative, drifting just below evens on a team that had "tournament favourite" energy.

On scorelines, every model that submitted a pick went with 2-1 — a perfectly sensible guess that the actual 5-1 made look rather conservative. All registered outcome hits, none exact. That's normal: exact scores are hard and 2-1 is almost always the statistically defensible call for a moderate favourite. The real story here is Grok's confidence paying off. Its betting-market lean pushed it furthest from a draw and furthest from a Sweden win, which happened to be the right direction to lean. No complaints.

Gemini's fingerprint — over-weighting league prestige — is worth noting here. Sweden's Allsvenskan pedigree presumably kept Gemini's draw probability slightly elevated (28%, the highest of any model). It didn't cost Gemini the outcome call, but in a game that finished 5-1 that extra draw weight looks soft. Claude's known home-advantage inflation didn't obviously fire here since Netherlands were already the favourite, but Claude's ceiling of 50% on a team that just won by four is arguably a symptom of the same conservatism.

Germany 2-1 Ivory Coast — Toronto

The cleaner match of the day, and the one where the scoreline picks earned their moment. GPT-5.4, Claude, and Gemini all called 2-1 exactly — and 2-1 is what happened. Three exact hits from three models in the same match is unusual enough to be worth celebrating, even if we keep the champagne corked: 2-1 is the commonest scoreline in football and models know this. Still, getting the direction and the margin right simultaneously is the hardest target in this exercise, and they deserve credit.

Grok was again the most bullish on Germany — 71%, a full 13 points above Claude and Gemini — and again posted the best Brier score (0.130). Betting markets clearly had Germany heavily favoured, and Grok followed suit. The irony is that Grok's own scoreline pick of 2-0 missed by one goal; the models that hedged slightly toward Ivory Coast scoring got the scoreline right. There's a lesson in there about confidence versus granularity, but one day's data doesn't make a rule. Gemini's away probability of just 17% for Ivory Coast was its lowest among the main models, consistent with its tendency to respect big European leagues — though here that bias and the actual result were pointing the same direction.

Ecuador 0-0 Curacao — Kansas City

Right. Let's talk about the disaster. Ecuador were given between 72% (Claude) and 86% (Grok) to win this match. Every single model and the naive average called a 2-0 Ecuador win. The actual result was a goalless draw. Every model was wrong on the outcome. Every model was wrong on the scoreline. This is the most instructive moment of the day.

The Brier scores are ugly across the board — anything above 1.0 is a rough day, and 1.5 is a rough week. Grok is the worst performer here precisely because it was most confident: 86% on a team that drew 0-0. Claude was least wrong, at 72% home, which translates into a Brier of 1.219 — the best of a bad lot, and consistent with Claude being the most cautious on heavy favourites. This is actually an argument for Claude's conservative style: when the shock result lands, you lose less. GPT-5.4 (78%) and the naive average also did relatively less damage. Gemini and Grok, both at the aggressive end, bled the most.

The collective miss is a finding, not an embarrassment. Ecuador vs Curacao was a match where the raw quality gap between the sides was enormous — Curacao are FIFA minnows, Ecuador a respectable South American side. The models did what any reasonable forecaster would do and loaded the probability onto the expected winner. Football, infuriatingly, didn't cooperate. This is a reminder that 80% doesn't mean 'certainty'; it means one match in five ends like this. The models weren't broken — they were unlucky in a statistically foreseeable way. The concern, if you want to find one, is that the draw probability across the field was squeezed to 11-17%. Draws in tight, low-stakes group matches between mismatched teams are under-priced by reputation-driven models. Something to watch.

Tunisia 0-4 Japan — Monterrey

Japan were favourites, Japan won convincingly, and the models all got the direction right — though again, nobody predicted the margin. All five models called 0-1; the actual 0-4 is, like the Dutch result, a reminder that models calibrate for likely scorelines rather than extremes.

The probabilities are unusually tightly clustered here — all models within a few points of each other. That consensus is sensible: Tunisia at home (Monterrey is not Tunis, so the 'home' tag is nominal) against a Japan side that qualified comfortably. Claude and Gemini share the best Brier score at roughly 0.23, both at 61% for Japan. Grok was the most cautious at 55%, which is interesting given its usual tendency to follow betting lines — perhaps the market was more circumspect about Tunisia than the other models expected. Gemini's 14% home probability for Tunisia is its lowest, consistent with its league-prestige weighting (Japan's J-League has climbed considerably in FIFA ranking terms). The low Tunisia home percentage from Gemini turned out well here, but it's worth flagging the mechanism: this isn't Gemini 'seeing' Tunisia's weaknesses, it's Gemini's structural preference for recognised top-flight nations.

Day summary

• Grok had its best day of the tournament so far — top Brier score in two of four matches, paid for being confident in the right direction. Ecuador cost it heavily, though.
• Claude finished second overall by being the least wrong in every match it missed — its conservatism is a genuine defensive tool, even if it leaves Brier points on the table when big favourites win.
• Gemini rode its Japan conviction well but was stung in Kansas City. Its aggressive home-win percentage for Ecuador (82%) is a symptom of over-weighting the reputation gap.
• GPT-5.4 was the most consistent across the slate — never the best, never the worst — which is its pattern. Two exact-score hits (Germany) are worth noting.
• The ensemble benefited from averaging out the extreme cases but couldn't escape the Ecuador collapse. Averaging bad priors just gets you average badness.
• Draws remain chronically under-priced. Ecuador-Curacao had draw probabilities of 11-17%; the draw happened. This is a systemic issue across the whole model field, not just one bad call.

What Sunday tests

Sunday's slate — check the fixtures — should stress a few known weaknesses. Any match involving a strong South American or African side against European opposition in a nominal 'home' slot will probe Claude's home-advantage inflation again. If there are matches between sides with significant FIFA ranking gaps, watch how tightly the draw probabilities are squeezed: after today's Ecuador lesson, that number deserves more scrutiny than the headline win probability. And if Grok continues to post the sharpest odds-aligned probabilities, we'll need more matches before deciding whether that's signal or Saturday's variance. Three from four on outcomes is a decent day. The scoreboard, though, keeps laughing at everyone's scoreline picks — and that's rather the point.