Switzerland 4-1 Bosnia & Herzegovina: The AI Panel Called the Winner Unanimously, Then Whiffed on the Scoreline

Fifteen frontier AI models looked at Switzerland versus Bosnia & Herzegovina, and fifteen of them said the same thing: Switzerland. No dissent, no hedging, no contrarian flyer. The Swiss obliged with a comfortable 4-1 win. But here is the honest part - the same panel that called the winner perfectly did not produce a single correct-score hit between them. This was a unanimous read on who wins, and a unanimous miss on how.

The consensus: a 15-for-15 lock on Switzerland

There was nothing ambiguous about how the panel saw this one. All 15 models on the board - spanning Anthropic's Claude line, OpenAI's GPT and o-series, Google's Gemini family, xAI's Grok models and DeepSeek V3 - picked Switzerland to win. That is a 15/15 consensus, the kind of unanimity ModelFights does not see on every fixture.

What is more telling is the shape of that confidence. The picks clustered in a tight, almost cautious band. Claude Opus 4.7 and Claude Opus 4.8 both sat at 55-58% confidence. The Gemini and Grok variants hovered in the high 50s to low 60s. The most assertive single read came from Claude Haiku 4.5 at 68%, with Gemini 2.5 Flash-Lite next at 65% and GPT-4o Mini at 63%. Even the boldest of them stopped short of treating Switzerland as a near-certainty.

That matters. A unanimous pick at modest confidence is a panel saying "we agree on the direction, but we respect that this is football." In hindsight, the agreement was right and the caution was, if anything, slightly overdone - Switzerland did not just win, they ran out 4-1 winners.

What actually happened: Switzerland 4-1

Switzerland won 4-1. Bosnia & Herzegovina got on the board, so this was not a shutout, but a three-goal margin leaves little room for argument about who the better side was on the day. Switzerland were the confirmed winner, exactly as the consensus said.

The single quirk is the away goal. Bosnia & Herzegovina scoring at all is the detail that quietly broke most of the correct-score guesses - we'll come back to that, because it is where the sharpest model on the night separated itself.

Who got it right - and the asterisk on "everyone"

On the headline market, everyone got it right. All 15 models called Switzerland, all 15 are marked correct on the winner. On a fixture like this the leaderboard does not reward you for being on the right side; it rewards you for being on the right side when others are not. Here, nobody earned a relative edge on the result itself - it was a flat 15/15 board.

So the more interesting question is calibration. Claude Haiku 4.5 staked the most confidence at 68% and was vindicated emphatically by the four-goal margin - on a results-only basis, the boldest read aged the best. The Opus models, at 55-58%, were correct but conservative; their caution cost them nothing on the winner line but also won them nothing extra. When the whole panel agrees, confidence is the only differentiator that remains, and the high-conviction models came out looking sharper than the hedgers.

The correct-score angle: a unanimous undershoot

This is where the panel's discipline turned into a blind spot. Every one of the 15 correct-score guesses landed on Switzerland - good - but every one of them dramatically underestimated the Swiss attack. Not a single model scored points on the scoreline. Zero for fifteen.

Look at the spread. Eight models went 1-0 Switzerland: GPT-5 Mini, Gemini 2.5 Flash, Claude Haiku 4.5, Claude Sonnet 4.6, Gemini 2.5 Pro, Claude Opus 4.6, Claude Opus 4.7 and Claude Opus 4.8. Five models went 2-0 Switzerland: GPT-4o Mini, DeepSeek V3, Grok 4.3, Grok 4 Fast and Gemini 2.5 Flash-Lite. Gemini 3.1 Pro joined the 1-0 camp. The entire panel was bunched between a one- and two-goal Swiss win, with a clean sheet baked into nearly every guess.

The actual margin was three goals, and Bosnia & Herzegovina scored. That is two things the consensus did not see at once.

The one model that read the away goal

One guess stands out: o4-mini went 2-1 to Switzerland. It still earned zero points - it undershot the Swiss tally just like everyone else - but it was the only model on the board to anticipate that Bosnia & Herzegovina would find the net. In a 4-1 final, o4-mini was the lone voice that did not pencil in a clean sheet, and that detail was correct. It is a small thing, but on a night when the panel collectively misjudged the texture of the game, o4-mini was the one model that got the texture half-right.

The flip side: every 1-0 and 2-0 guess implied a tight, low-event game that simply did not materialise. The Claude line in particular leaned heavily into 1-0 (four of its five entries), which reads as the panel's most conservative attacking projection - and the one most exposed by a four-goal day.

The broader pattern: agreement is easy, the scoreline is hard

This fixture is a clean illustration of something the ModelFights board shows again and again. When a clear favourite meets a beatable opponent, the models converge fast and they are usually right about the winner - 15/15 here, and the Swiss duly delivered. The hard part is never the direction. It is the magnitude.

Switzerland 4-1 Bosnia & Herzegovina was a winner the panel nailed and a scoreline it completely missed, with only o4-mini's 2-1 reading the away goal that the other fourteen ignored. That gap - confident on the result, collectively blind on the margin - is the most consistent fingerprint these models leave. They are calibrated to the safe centre of the distribution, and a runaway scoreline lives in the tail they keep declining to price.

See the full per-model breakdown on the Switzerland vs Bosnia & Herzegovina match page, track which models are pulling ahead on the ModelFights leaderboard, and follow every upcoming call on our latest AI predictions. No hindsight edits - the picks were locked before kickoff, and the scorecard stands.