Which AI is best at predicting Premier League results?

A Premier League season is 380 matches. Slice it by model, by month, or by "recent form", and you are ranking on samples of a few dozen — where luck comfortably outweighs skill.

The arithmetic is unforgiving. Honest three-way result accuracy in football sits inside a band a few percentage points wide between serious systems, so separating two of them takes hundreds of matches, not a matchweek. Across thirty Premier League fixtures a model can finish top by getting two coin-flips right; next month a different one does.

That is why ScoreGPT flags low-sample windows on its own pages instead of dressing them up as standings. A per-competition row built on too few graded matches is marked as such, because the alternative is inviting you to read a ranking out of noise. The general version of this argument is on Which AI is best at football predictions?

There are honest reasons to expect some variation — and none of them survive a small sample. Leagues differ in ways a model can read well or badly. The Premier League's fixture congestion and deep squads make rotation genuinely hard to call. Its competitive spread means bottom-half sides beat title challengers every season. Its coverage is so dense that team news is abundant, and often late.

So a model that weighs rest and rotation heavily might really do relatively better here than in a league where the top two win almost everything. The catch is that you cannot detect an effect that size inside one season of one league — it is smaller than the noise around it. Treat per-league differences as a hypothesis worth watching across years, not a ranking to act on this week.

Two public surfaces carry it, and both are computed live rather than typed into copy.

• The Premier League hub — every covered fixture with the consensus call and each model's lean before kickoff, and the running graded record for the competition below it, hits and misses in the same list.
• Each model's record page — a per-competition table for the last 30 days showing graded count, result accuracy and exact-score accuracy, with a low-sample flag whenever the count is too thin to read as a ranking.

Deliberately, no Premier League accuracy percentage is written into this page. A figure typed into static copy is stale by the next matchweek, and on a single-league sample it swings far more than the underlying quality does. The live pages are the claim; this page is only the map to them.

The Premier League season kicks off in mid-August, and the record for the current model panel starts from there. ScoreGPT's five-model roster — GPT-5.6 (OpenAI), Claude Opus 4.8 (Anthropic), Grok 4.5 (xAI), GLM-5.2 (Z.ai) and Kimi K3 (Moonshot) — has been running since 20 July 2026, so the Premier League slice of its record is a new book opened at the first matchweek, not a continuation of an older roster's numbers.

That is on purpose. When a lab ships a stronger successor the model changes, and graded history under the old version does not silently transfer to the new one. It also means the honest thing to say in August is that the sample is small — and the pages say exactly that, rather than presenting a handful of matchweeks as a league table of AIs.

Every covered Premier League fixture is read independently by all five models, combined into one call, and checked in public afterwards. For each match one file is assembled the same way: recent form and who it came against, head-to-head history, injuries and suspensions, rest days and midweek European travel, what the match means for each side, and market context. Predictions land daily around 07:30 UTC as fixtures enter the analysis window.

On the free web pages you get the consensus result, the predicted scoreline and each model's lean; the app adds every model's own scoreline, confidence and reasoning, with 3 AI predictions a week free and Pro removing the cap. After full time everything is public — every model's pick marked right or wrong, and it stays on the match page. How the pipeline works end to end is on the methodology page.