Five Matches, One Shock, and a Mexican Masterclass

Wednesday, 24 June served up five matches across Groups A, B, and C — all kicking off in the same two windows. Four of them went to the expected side. The fifth, South Africa versus South Korea in Monterrey, did not. More on that shortly, because it deserves its own post-mortem.

Bosnia and Herzegovina 3–1 Qatar (Seattle)

The models were in comfortable agreement here: Bosnia were favourites, Qatar were there to be beaten. Probabilities ranged from Gemini's rather modest 58% for a home win up to Grok's bullish 71%. The home win duly arrived, so all models got the outcome right. The interesting variance is in calibration: Grok's higher conviction was rewarded with the best Brier score on the day (0.13), while Gemini's relative hesitancy left it as the worst performer of the group (Brier 0.27). Note that Bosnia's home advantage here is a neutral venue — Lumen Field in Seattle — so Claude's known tendency to over-price home advantage didn't have a structural opportunity to inflate its number, and indeed Claude and GPT-5.4 both sat at 62%, a touch below the ensemble.

On scorelines, all models picked 2–1 or 2–0 for Bosnia — sensible, compact wins. The actual 3–1 was a touch more emphatic, but all had the direction correct. No exact hits, which is entirely normal, and a miss-by-one-goal scoreline isn't really a miss worth dwelling on.

Switzerland 2–1 Canada (Vancouver)

This was the most genuinely uncertain match on the slate, and the models reflected that honestly — perhaps a little too honestly, since they largely hedged their way into indecision. GPT-5.4, Claude, and Gemini all had the three outcomes within a few percentage points of each other, essentially shrugging. Grok was the odd one out, nudging Switzerland to 47% and Canada down to 23%, which turned out to be the shrewder position. Switzerland's win rewarded Grok with the best Brier score here (0.42) while Claude and Gemini trailed at 0.58.

This is a match where Grok's tendency to lean on betting-market odds may have helped it — the markets likely had Switzerland as mild favourites, and Grok tracked that. Claude and Gemini both sat at 38% for Switzerland, their most cautious reading of the three. All models picked 1–0 as their scoreline; the actual 2–1 was correct in direction and margin but not in exact score. These were the right instincts, just slightly underpowered.

Morocco 4–2 Haiti (Atlanta)

Routine on the outcome, interesting on the detail. Morocco were heavy favourites across the board — Grok most confident at 84%, Claude most conservative at 75%. The home side won, as expected, and Grok's conviction again paid off with the best Brier score (0.04 — genuinely excellent for a 4–2 thriller). Claude's 75% looks modest in hindsight but is arguably more honest given the uncertainty around a 4–2 final score; a 84% implies a margin of comfort that a two-goal concession somewhat complicates.

Haiti scoring twice is a reminder that even well-calibrated high-probability predictions can hide surprises within the outcome. All models picked 2–0 Morocco; the actual 4–2 shows everyone underestimated the goal volume. No exact hits, but every model nailed the direction with reasonable confidence.

Scotland 0–3 Brazil (Miami Gardens)

The field got this one right, and got it right with some conviction. Brazil were overwhelming favourites — Gemini highest at 76%, Grok lowest of the group at 69% (a rare moment of Grok underplaying a gap between reputation and reality). The 0–3 scoreline validated the stronger probabilities, with Gemini picking up the best Brier score (0.09) and Grok the worst of the group (0.15). This is one of those matches where Claude's known reputation-versus-form bias didn't really manifest — Brazil's form and reputation aligned, so there was nowhere for the bias to hide.

Every model picked 0–2 Brazil as their scoreline; the actual 0–3 was more emphatic but directionally identical. A clean sweep of outcome hits with no exact hits is exactly what you'd expect from a match this lopsided.

Czechia 0–3 Mexico (Estadio Azteca)

Mexico won comfortably at the Azteca, which the models half-expected — Mexico were favourites across the board, but at fairly modest margins. Claude led with 50% for Mexico, GPT-5.4, Gemini, and Grok clustered in the 43–47% range. The 0–3 scoreline was emphatic in a way that 47% doesn't fully capture; models got the direction right but were essentially treating this as a coin-flip-ish contest when Mexico turned it into a stroll.

This is a venue where context matters enormously. The Azteca carries enormous weight for Mexican football, crowd and altitude included. None of the models appeared to price that atmosphere heavily — and yet the result arguably reflected it. Claude's 50% was the sharpest read; Grok and Gemini's ~43–44% look slightly too cautious in hindsight, which is interesting given this is exactly the kind of match where Gemini's league-prestige weighting might have been expected to boost Mexico (a CONCACAF side playing at home in a World Cup) rather than restrain it.

South Africa 1–0 South Korea (Monterrey)

Here it is. Every model picked South Korea. Every single one. Probabilities for a South Korean win ranged from 52% (Grok) to 60% (Claude), with the remainder spread across draw and a South Africa win that nobody gave more than 21% to. South Africa won 1–0. All models were wrong on the outcome. All scoreline picks were wrong on outcome too — every model locked in 0–1 to South Korea.

The Brier scores here are damning — anything above 1.0 is genuinely bad probabilistic prediction, and Claude's 1.12 is the worst of the tournament so far for a single match. Claude gave South Africa just 16%, its lowest home-team probability of the day, which is striking given its known tendency to over-price home advantage. In this case, South Korea's reputation simply overpowered that bias entirely, and the result punished it hardest. Grok, by virtue of being least certain about South Korea (52% rather than 55–60%), took the least damage — still wrong, but least wrong.

This is a finding, not an embarrassment. A 16–21% probability outcome will happen roughly one in five to six times. That is not an indictment of the models' logic; South Korea were the stronger side on paper. But collectively giving South Africa no more than a one-in-five chance — and in Claude's case one-in-six — means the models shared a blind spot. Whether that's over-reliance on FIFA rankings, insufficient weight given to South Africa's African Cup experience or South Korean form concerns, or simply the fact that narrow margins in knockout-adjacent group games resist prediction, is worth investigating as the tournament data accumulates.

Day Summary: Who Performed Best?

Across the five matches, Grok had the best day: it topped or co-topped the leaderboard on Bosnia, Switzerland, and Morocco, and suffered least on South Africa. Its habit of anchoring to betting-market odds served it well on a day when the markets were broadly right (minus Bafana Bafana). Claude had the worst day, anchoring lowest on three matches it got right but paying the steepest price on South Africa. Gemini bounced — brilliant on Brazil, poor on Bosnia. GPT-5.4 was consistent but unspectacular.

• Grok: Best single-match Brier scores on Bosnia (0.130) and Morocco (0.042); least wrong on South Africa (0.967). Best overall day.
• Gemini: Best on Brazil (0.090) thanks to highest confidence in the favourite; dragged down by Bosnia (0.268) and South Africa (1.054).
• GPT-5.4: Steady mid-table performer. No standout wins, no catastrophic losses beyond South Africa.
• Claude: Paid the highest price for the South Africa shock (Brier 1.123). The over-reliance on reputation over form is the clearest fingerprint on display today.
• Ensemble: Did what ensembles do — absorbed the extremes, never first, never last, consistent mediocrity in the best sense.

What Thursday's Slate Will Test

Tomorrow brings more group-stage football with sides whose qualification positions are becoming clearer — which means we'll see matches where one team has something to play for and another doesn't. That's precisely the context where reputation-based models struggle most: a ranked side with nothing at stake versus a lower-ranked side desperate for points. Watch whether the models adjust for motivation, or whether they keep defaulting to quality metrics alone. South Africa today was a case study in motivation mattering. It won't be the last.