Grok nails Brazil, everyone flunks Germany: Monday's World Cup AI report

Monday's Round of 32 slate gave us one comfortable home win, one almighty shock, and one competitive stalemate — a useful stress test across the full range of outcomes. The models went 1-for-3 on results, which sounds bad and largely is, though the manner of the failures is more instructive than the headline number.

Match 1 — Brazil 2-1 Japan

The most straightforward match of the day, and the one where the models did their best work. Brazil won as the favourite; every model called the outcome correctly. The interesting question is *how confidently* they called it — and here Grok stands out.

Grok's 0.67 on Brazil was the most decisive call and earned comfortably the best Brier score (0.169) by some distance. Its known fingerprint — leaning hard on betting-market odds — served it well here; the markets had Brazil at roughly 0.56 implied probability, and Grok pushed beyond that, which turned out to be the right direction. The ensemble also performed creditably, its aggregation pulling it clear of most individual models.

Claude and GPT-5.4 were joint-lowest at 0.52 on Brazil. Claude's known tendency to over-price home advantage is not the story here — if anything, 0.52 looks *under*-confident for a side of Brazil's standing. That said, Japan are no pushover in this era, and 0.52 is a defensible number; it's just that Grok's conviction was better rewarded on the day. On scorelines, every model picked 1-0 for Brazil (Grok went 2-0), so all got the direction right; none landed the exact 2-1, which is routine at this level of granularity — no shame in a near-miss when you've called the winner.

Match 2 — Germany 1-1 Paraguay

This is where Monday got uncomfortable. Germany drew with Paraguay, a result the models did not so much fail to predict as actively resist predicting. Every single model — and the market — loaded the vast majority of probability onto a German win, leaving draw probabilities in the 17–23% range. When the draw duly arrived, everyone was punished.

Claude and GPT-5.5 were the *least wrong*, though that is a low bar when every model posted a Brier score north of 1.0. Claude's 0.62 on Germany was the most cautious home call — interesting, given that Claude supposedly over-prices home advantage, but Germany are not a neutral-venue home side here, so perhaps the fingerprint is more situational than blanket. GPT-5.5's 0.23 draw probability was the highest on the panel, which is why it leads the damage-limitation table.

Grok was the most severely punished: 0.76 on Germany, just 0.17 on the draw, Brier of 1.271. Its reliance on betting-market signals backfired badly — the market itself was at 0.716 for Germany, and Grok went further still. The market's Brier of 1.190 was itself dire, which tells you this was a genuinely surprising result rather than one the models should have seen coming with better calibration. Even so, the collective under-pricing of draws against lower-ranked opponents is a pattern worth filing. All six scoreline picks were 2-0 to Germany; all six were wrong on outcome, not just the exact score.

Match 3 — Netherlands 1-1 Morocco

The late kick-off produced a second consecutive draw, and a second consecutive collective model failure — though a softer one. The Netherlands–Morocco odds were much tighter than Germany–Paraguay, with the panel giving Netherlands 38–47% and Morocco 23–30%, so the draw was at least treated as a live outcome even if no model made it the modal call.

Gemini wins this one — its 0.34 draw probability was the highest on the panel and its 0.38 Netherlands call the most conservative, yielding the best Brier score (0.658) by a clear margin. Notably, Gemini's fingerprint is over-weighting league prestige, but here it seems to have processed Morocco's strong recent tournament pedigree with enough respect to trim the Dutch win probability. Whether that's the fingerprint working correctly or coincidentally arriving at the right number is hard to say from one data point, but it's worth watching.

Grok again leaned towards the favourite (0.47 Netherlands) and paid a moderate penalty. GPT-5.4 was the most wrong model in this match. Claude and GPT-5.5 posted identical distributions (0.40 / 0.30 / 0.30) — a perfectly balanced sheet that, while technically wrong, at least gave the draw its fair share of probability. Every scoreline pick was 1-0 to Netherlands; all six were wrong on outcome. Morocco's parity with a major European side continues to confound the field.

Day summary

• Grok had its best individual performance of the day on Brazil–Japan (Brier 0.169) and its worst on Germany–Paraguay (1.271). High conviction is a double-edged blade.
• Gemini was the most accurate model on Netherlands–Morocco, vindicated by its relatively high draw probability. Its Germany call was mid-table bad.
• Claude was the joint-least-wrong on Germany–Paraguay (with GPT-5.5), consistent with its more cautious win probabilities for the home side — though 0.62 is still a heavy lean.
• GPT-5.5 placed best or near-best in two of three matches; a quietly solid day relative to peers.
• GPT-5.4 had no standout result and finished at or near the bottom in both draws.
• The ensemble beat most individuals on Brazil–Japan but was mid-to-bottom on both draws, showing aggregation helps with clear favourites but doesn't rescue collective blind spots.
• Two consecutive draws involving a clear favourite underscore the tournament-wide pattern: models undervalue the draw in knockout-adjacent pressure contexts.

What Tuesday's slate might test

After two high-profile draws today, Tuesday's matches will be an immediate test of whether the models recalibrate draw probabilities upward or persist with the same favourite-leaning distributions. If the next slate includes any match pairing a top-ten-ranked side against a tournament dark horse, watch whether Gemini's relative conservatism or Grok's high conviction is better rewarded. The Germany result in particular should put every model's Germany-adjacent confidence on notice: reputational weight is not the same as match-day certainty.