Portugal 1–1 DR Congo: 15 AI Models, 105 Predictions, and Only 2 Right

Same brief, same scoreboard, public grading — no cherry-picking. Fifteen of the best models on earth made 105 calls on this game. Two of them landed.

There are upsets, and then there is the kind of result that quietly dismantles an entire panel of predictions. Portugal 1–1 DR Congo at the 2026 World Cup was the latter. ModelFights handed the same pre-kickoff brief to fifteen frontier AI models, locked every pick with a public prompt hash before a ball was kicked, and graded the lot the moment the final whistle blew. Across seven betting markets that came to 105 individual predictions — and when the dust settled, just two of them were correct.

This is the full, graded story: who picked what, how confident they were, what the market thought, and exactly how a single goal from DR Congo turned a near-unanimous Portugal scriptline into one of the heaviest collective misses of the tournament.

TL;DR

• All 15 models backed Portugal to win, at an average of 79% confidence — and the market agreed, with Portugal priced around 1.33 (~75% implied).
• That includes the entire flagship tier: GPT-5, Claude Opus 4.8, Gemini 3.1 Pro and Grok 4.3.
• The panel also leaned Portugal –1.5, Over 2.5 goals, both-teams-to-score "No", a Portugal HT/FT double, and a 2–0 scoreline.
• Final score: 1–1. Of 105 graded predictions across seven markets, only 2 won — the two models that broke from the pack on both-teams-to-score.

The matchup: a genuine favourite, not a coin flip

Let's be clear about the setup, because it matters. This was not a 50/50 the models talked themselves into. Portugal arrived as one of the marquee names of the World Cup and a heavy favourite; DR Congo as the underdog with everything to gain. The bookmakers had Portugal at roughly 1.33 on the moneyline — an implied probability around 75% — and the AI panel, on average, was a touch more bullish than that. When the models and the market both lean the same way this hard, you are looking at a result the entire prediction industry expected. Football had other ideas.

Match winner: a clean sweep for Portugal

Every single model picked Portugal. The confidence band was tight and high — fourteen of fifteen between 74% and 85% — with no genuine dissent anywhere on the board:

At ~79% average confidence against a market price implying ~75%, the panel saw no edge worth fading and no reason to hedge. On the match-winner market, that conviction returned 0 wins from 15.

Even the flagships were all-in

It is tempting to write off a missed call as a small, cheap model getting it wrong. That excuse does not survive contact with this game. The premium reasoning tier — Claude Opus 4.8 and Opus 4.7, GPT-5, Gemini 3.1 Pro and Grok 4.3, the most capable models you can run — all backed Portugal. If anything, the flagships were marginally more measured (Opus 4.8 and 4.7 were the most cautious at 74%) than the budget models (GPT-4o Mini and Gemini Flash-Lite topped the board at 85%). The direction, though, was unanimous. This was not a failure of horsepower. It was football refusing to read the team-sheet.

Market by market: the panel leaned Portugal everywhere

The moneyline was only the headline. Across all seven markets, the models painted the same picture — a comfortable, front-foot Portugal performance — and the 1–1 draw took an axe to nearly every line.

Asian handicap — Portugal –1.5 (10 of 15)

Ten of fifteen took Portugal to cover a 1.5-goal handicap, i.e. to win by two or more, at around 1.95. A draw is about as far from "win by two" as it gets. Graded: 0 wins, 11 losses, 4 pushes.

Spread –1 — Portugal (8 of 15)

The straight –1 spread told the same story: Portugal to win by a clear margin. Result: 0 wins, 8 losses, 7 pushes.

Over / Under 2.5 goals — Over (12 of 15)

Twelve of fifteen expected Portugal's attack to flow and took the Over, around 1.68. The game produced exactly two goals — under the line. The overs lost across the board.

Both teams to score — No (13 of 15)

Here is where the story has its only twist. Thirteen models said DR Congo would be shut out. They were not — DR Congo scored, and "No" lost 13 times. But the two models that broke ranks and called "Yes" were the only two correct predictions on the entire 105-pick board. On a day when the consensus was wrong about everything, the contrarians on one market were the whole story.

Half-time / Full-time — Portugal / Portugal (9 of 15)

Nine models called Portugal to lead at the break and win at the whistle. A 1–1 finish kills the double outright. Graded: 0 wins, 15 losses.

Correct score — 2–0 Portugal (8 of 15)

The most-backed exact scoreline was 2–0 Portugal — fittingly, a clean, comfortable home win. Correct-score is the highest-variance market on the board (average confidence here was just 22%, the models' honest way of saying "this is a long shot"), and 1–1 was not on anyone's shortlist. Graded: 0 wins, 15 losses.

The scorecard: 2 of 105

Tally it up and the scale of the miss is hard to overstate. Fifteen models, seven markets, 105 graded predictions — and exactly two cashed. Not two markets; two individual picks, both on the same "both teams to score: yes" call. Every moneyline, every handicap, every spread, every Over, every HT/FT, every correct score: wrong. It is, by raw hit rate, the single worst collective performance the panel has put up at this World Cup.

What actually happened

DR Congo did not come to admire Portugal. They defended with discipline, took their moment when it came, and walked off with a point that the data said they had almost no business taking. There is no clever spin to apply: the favourite was held, the goals the models expected never arrived in the quantity they expected, and a single DR Congo strike flipped the goals and both-teams-to-score markets on their head at once.

Why it matters

This is the cleanest illustration yet of a pattern we have tracked all tournament: AI models are excellent at ranking teams and helpless at conjuring a draw. When a favourite is genuinely better, the panel says so with conviction — and most of the time it is right. But the World Cup group stage keeps producing the one outcome the match-winner market cannot express, and a confident, unanimous read on a strong team is still not a certainty. Portugal 1–1 DR Congo now sits alongside a growing list of group-stage stalemates that have quietly gutted the AIs' record — the full pattern is in our blowouts vs draws breakdown.

The consolation: these models learn

Here is the part worth remembering. Every graded result — including this one — becomes part of the public record these models are measured against. A miss is not just a miss; it is a data point. The interesting question is never whether an AI got a single match wrong (they all will, repeatedly), but whether the panel adjusts — whether its calibration tightens, whether it learns to price in the chaos of a one-off knockout-style game. We will be watching the next Portugal fixture with the picks locked and the receipts public, exactly as we did here. That is the whole point of doing this in the open.

How ModelFights works

Every model gets the same brief. Every pick is locked before kickoff and stamped with a public prompt hash, so nothing can be quietly rewritten after the result. Every market is graded in the open — the misses stay on the record next to the hits, because a confident, unanimous call that ages badly is exactly the kind of thing most tipster sites memory-hole. We keep the receipt.

Where to follow it live

The full, immutable pick record for this match — all fifteen models, all seven markets, every confidence figure and the odds each model saw — is on the Portugal vs DR Congo prediction page. To see which model is actually beating the closing line across the whole tournament, check the AI leaderboard, and browse today's locked picks on the predictions board.

FAQ

What was the score of Portugal vs DR Congo?

Portugal drew 1–1 with DR Congo at the 2026 World Cup.

Who did the AI models predict to win Portugal vs DR Congo?

All 15 frontier models — ChatGPT/GPT-5, Claude, Gemini, Grok and DeepSeek — predicted Portugal to win, at an average of 79% confidence. The match drew 1–1, so every match-winner pick lost.

Did any AI predict Portugal vs DR Congo correctly?

Not on the winner. Of 105 graded predictions across seven markets, only two were correct — the two models that called "both teams to score: yes." Everything else missed.

Which AI was most confident Portugal would win?

GPT-4o Mini, Gemini 2.5 Flash-Lite and Gemini 3.1 Pro led at 85%. The most cautious — still backing Portugal — were Claude Opus 4.8 and Opus 4.7 at 74%.

Were the premium AI models any better than the cheap ones?

No. The flagship reasoning models (Claude Opus 4.8, GPT-5, Gemini 3.1 Pro, Grok 4.3) all backed Portugal in the same confidence band as the budget models. On this match, model size made no difference.

Final word

Portugal should have won. Fifteen of the world's best models said so, the market said so, and for long stretches the game probably said so too. But the World Cup does not pay out for "should have", and a 1–1 draw is the quiet, recurring reminder that a confident AI is not a certain one. Two picks out of 105 is a brutal scoreline for a panel this strong — and the most honest thing on this page.