Every Premier League gameweek, we ask ChatGPT, Claude and Gemini to independently predict the exact score of every match – then benchmark them against a fourth competitor built from sportsbook market prices.

The experiment is designed to keep the playing field as level as possible.

Each AI receives the same fixtures, the same prediction deadline and the same core competition prompt. They’re allowed to independently research current football information, but they are not allowed to use bookmaker odds, betting tips, prediction websites or other AI predictions to inform their selections.

The question we want to answer over a full Premier League season is simple:

Who predicts football best – ChatGPT, Claude, Gemini or the sportsbook-price benchmark?

The Competitors

ChatGPT – OpenAI GPT-5.6

ChatGPT independently researches each fixture using current football information before producing its predicted score, 1X2 probabilities, expected goals and supporting reasoning.

Claude – Claude Sonnet 5

Claude follows the same competition instructions and prediction deadline while performing its own independent research and analysis.

Gemini – Google Gemini 3.1 Pro

Gemini independently researches each fixture and generates its own probability estimates, expected goals and exact-score prediction.

Bzeebet Market Model

Unlike the three AI competitors, the Bzeebet Market Model does not use a language model.

It is a rules-based benchmark built from Playzee sportsbook prices captured at the prediction deadline.

First, the shortest Playzee 1X2 price determines the predicted result:

Home win
Draw
Away win

Bzeebet then considers only Correct Score selections that are consistent with that result and selects the shortest-priced compatible scoreline.

For example, if Draw is the shortest 1X2 price, only scores such as 0-0, 1-1 or 2-2 can be selected. A home or away winning score can never be used, even if that individual Correct Score price is shorter.

This gives the three AI models a consistent sportsbook-price benchmark rather than simply adding another prediction model.

Odds are sourced from Playzee, a Bzeebet affiliate partner, and archived at capture time. Playzee provides the sportsbook prices used by the model but has no role in the AI predictions, competition scoring or results.

Same Rules. Different Brains.

he three AI models are deliberately kept separate from each other and from sportsbook market data.

They do not see:

Bookmaker odds or market probabilities
Betting tips
Prediction or tipster websites
The other AI models’ predictions
The Bzeebet Market Model prediction

They are encouraged to independently research factual information including:

Recent results and form
Injuries and suspensions
Confirmed team news
Fixture congestion
Home and away performance
Managerial changes
Current-season statistics
Relevant previous-season context

Where possible, research is based on authoritative sources such as official Premier League, club and competition websites.

Because each model conducts its own independent research, the three AIs can receive identical instructions and still reach very different conclusions.

Prediction Deadline

Predictions are normally generated and locked approximately 24 hours before the first match of each Premier League gameweek.

Once locked, a prediction cannot be changed because of later:

Team news
Injuries
Suspensions
Starting line-ups
Market movement
Other models’ predictions

The Bzeebet Market Model uses Playzee prices captured around the same prediction deadline.

All locked predictions, probabilities and associated metadata are retained as historical records and are not retrospectively edited after matches are played.

Scoring System

All four competitors are scored under exactly the same rules.

Correct match result – 3 points

For example, if the prediction is:

Arsenal 2-1 Chelsea

and Arsenal win by any scoreline, the competitor receives 3 points.

Exact score – 5 points total

If the match finishes exactly:

Arsenal 2-1 Chelsea

the competitor receives 5 points total.

The scoring is therefore:

Exact score = 5 points, not 3 + 5.

This means competitors must balance selecting the most likely match result with identifying the most plausible exact score.

Prompt Transparency

We want the Premier League AI Battle to be a transparent and reproducible experiment rather than simply publishing unexplained AI predictions.

All three AI competitors use the same core competition prompt.

Current production prompt: v5.1

SHA-256:

3d40561c77e1691cc7d823860d674b13e08a3d88e0d531d497d04100e265e1f4

The cryptographic hash allows us to record exactly which prompt version was used for each prediction run and makes any future methodology changes visible.

What Each AI Must Produce

For every Premier League fixture, each AI returns:

Predicted home score
Predicted away score
Predicted 1X2 result
Home-win probability
Draw probability
Away-win probability
Probability of the selected exact score
Expected goals for each team
Short reasoning
Research notes
Research sources

The Home, Draw and Away probabilities must total 100%.

The selected exact score must also be logically consistent with the predicted 1X2 result.

View the Full Production Prompt – v5.1

T

Methodology note: The same core competition prompt is supplied to ChatGPT, Claude and Gemini. Provider-specific model infrastructure, search systems and internal generation behaviour may differ.

[▼ View Full Production Prompt v5.1]

Research Quality Controls

The prediction system includes automated checks designed to keep the comparison consistent.

Research sources used by the AI models are reviewed to prevent reliance on prohibited material such as:

Betting odds
Betting tips
Correct-score tips
Match-prediction websites
Other AI predictions

The system also checks prediction output for issues such as:

Missing fixtures
Duplicate fixtures
Probabilities that do not total 100%
A predicted score that contradicts the selected 1X2 result
Invalid structured output

Some non-critical research issues may be recorded as quality-control warnings rather than invalidating an entire gameweek.

Why Include a Sportsbook Market Benchmark?

AI models analyse football in very different ways, while sportsbook prices incorporate a large amount of information about expected match outcomes.

That makes sportsbook pricing a useful independent benchmark.

If an AI model consistently outperforms the sportsbook-price benchmark across hundreds of Premier League predictions, that becomes considerably more interesting than simply showing that one AI scored more points than another AI.

Our primary measurement is:

Prediction accuracy – which competitor earns the most points?

As the historical dataset grows, we can also analyse how each model’s selections would theoretically have performed using the Playzee odds available when predictions were locked.

Historical Odds & Future Simulation

Bzeebet archives the Playzee prices captured when the Market Model is locked.

The same analysis can be performed for:

ChatGPT
Claude
Gemini
Bzeebet Market Model

We will be able to compare metrics including:

Number of selections
Strike rate
Total theoretical stake
Theoretical returns
Profit or loss
ROI

These simulations are retrospective analysis only and are not betting recommendations.

Editorial Disclaimer

The Premier League AI Battle is an editorial experiment designed for information, analysis and entertainment.

AI-generated football predictions can be wrong, and historical performance does not indicate future results.

None of the predictions published by ChatGPT, Claude, Gemini or the Bzeebet Market  Model should be considered betting advice or a recommendation to stake money.