We give 10 of the world's leading AI models the same match files and the same instructions. We track every bet, before kick-off and live. Who is best? And is any of it more than luck?
Profit at £10 a bet, on each strategy's most recent bets
Pre-match and in-play are ranked apart. The same AI models bet both ways, on the same matches with the same data, at three levels of risk. Reading a game before kick-off and during it are different skills, so the results are never mixed.
Bets placed before kick-off, from form, team news and prices.
Bets placed while the match is on, reading the live score and moving odds.
Picks land here the moment a model makes them, and flip to won or lost when the match ends. This page refreshes itself, so leave it open.
Newest first, across every model and competition.
Matches with picks still open. Results land as soon as the final whistle goes.
54 days in, 108,018 picks settled. These headlines rewrite themselves as results come in.
Profit if you had put £10 on each model's last 50 picks. Pre-match, the models get the same match data and bet before kick-off. Tap a name to follow it.
Each line adds up the profit of the model's last 50 settled pre-match bets (all three styles), oldest to newest, across all its strategies. A model with fewer bets joins the line late. Luck zone: where a picker with no skill at all would land 9 times out of 10 with the same number and kind of bets. It slopes down because bookmaker prices carry a margin (about 5% per pick).
Every model's picks pooled together: 108,018 settled bets. Where does AI as a whole find an edge, and where does it give its money away? Each answer says whether it's a real pattern yet or still too close to call.
Best: in-play (−5.1% per pick). Weakest: before kick-off (−5.8%). Even the best is behind a random picker so far.
Best: national teams (−3.2% per pick). Weakest: top-flight clubs (−6.2%). That beats a random picker, but isn't profit yet.
Best: handicap (−4.0% per pick). Weakest: match result (−10.1%). That beats a random picker, but isn't profit yet.
Best: favourites (−0.9% per pick). Weakest: long shots (−16.7%). That beats a random picker, but isn't profit yet.
Best: conservative (−1.6% per pick). Weakest: aggressive (−10.8%). That beats a random picker, but isn't profit yet.
A few dozen bets can make anyone look like a genius. So we ask a simple question of every model: could a picker with no skill at all, betting at the same odds, have done this well?
Compared with a picker who has no skill at all. Further right means harder to explain by chance; the grey middle is where pure luck lives.
Each dot is a picker with no skill: 6116 random bets at the same kind of real bookmaker prices the models took. The pins are the real models.
2 models are doing worse than 95% of random pickers so far. Nobody is beating the bookmaker's prices in this competition yet.
Every model bets at three risk levels: conservative, balanced and aggressive. Same data, same matches. Only the attitude to risk changes.
Green made money, red lost it. Stronger colour means a bigger swing. ★ marks the best model in each column.
Across every model, conservative has been the best style so far (−1.6% per pick). Aggressive was the worst of the three for 9 of 10 models.
Average odds taken in each style. A longer bar means the model changed its behaviour more when told to take risks.
Before kick-off a model has time, stats and team news. In-play it has to read a match as it happens and move before the price does. The same model can be a very different bettor in each.
Pre-match covers every strategy that bets before kick-off; in-play covers every strategy that bets while the match is on. Each model runs several strategies on each side, sometimes with more than one version, e.g. Kimi runs kimi-k2.6 in-play.
The same model can be sharp in one competition and lost in another. Here's return per pick for every model in the competitions with the most bets.
| 🇺🇸 USA Major League Soccer | 🇦🇷 Argentina Liga Profesional | 🇧🇷 Brazil Serie B | 🌍 UEFA Europa League | 🇧🇷 Brazil Serie A | 🇮🇹 Italy Serie A | 🇲🇽 Mexico Primera Division | 🏴 England League Championship | Best for it | |
|---|---|---|---|---|---|---|---|---|---|
| All AI | −10.0%4225 picks | −4.8%4187 picks | −0.9%3234 picks | +2.4%3049 picks | −9.7%3033 picks | −13.2%2771 picks | −4.0%2703 picks | −4.8%2698 picks | |
| Claude | −0.8%318 | −7.6%360 | −12.7%329 | +1.8%141 | −16.2%230 | −8.4%178 | −11.9%191 | −24.1%208 | UEFA Europa League |
| DeepSeek | −8.2%679 | −10.1%620 | −0.6%487 | +14.9%437 | −11.4%459 | −15.7%421 | −3.5%369 | −3.7%455 | UEFA Europa League |
| Gemini | −12.1%683 | −1.8%615 | −7.7%498 | +4.0%474 | −14.7%448 | −11.0%431 | +3.0%398 | −4.5%459 | UEFA Europa League |
| GLM | −9.9%561 | −7.0%524 | −1.0%434 | −2.8%415 | −6.4%401 | −13.9%370 | −6.2%342 | +0.9%353 | England League Championship |
| GPT | −11.9%640 | −4.5%595 | +8.9%490 | +11.6%439 | −9.4%398 | −21.5%378 | −1.5%391 | −3.4%381 | UEFA Europa League |
| Grok | −7.0%354 | −2.2%356 | +12.8%303 | −8.3%276 | −3.0%239 | −28.1%221 | −5.3%243 | +4.5%231 | Brazil Serie B |
| Kimi | −5.0%206 | −1.1%233 | +2.2%147 | +3.6%169 | +6.0%177 | −5.8%136 | −6.6%151 | +2.6%114 | Brazil Serie A |
| MiMo | −11.3%238 | −0.3%278 | +5.1%157 | −14.7%214 | +1.7%216 | −1.7%198 | −3.4%193 | −9.5%165 | Brazil Serie B |
| Mistral | −13.7%298 | −9.1%317 | −3.9%204 | +1.1%246 | −27.2%252 | −3.7%233 | −3.3%219 | −10.6%175 | UEFA Europa League |
| Qwen | −19.7%248 | +1.6%289 | −13.7%185 | −3.4%238 | −6.3%213 | −11.3%205 | −9.6%206 | −6.9%157 | Argentina Liga Profesional |
Cells with fewer than 8 picks show only the count. Country comes from the match data; continents mean national-team competitions.
Before kick-off every model picks from the same menu of bets. So when one keeps choosing handicaps and another keeps backing goals, that's the model's own taste showing.
Share of each model's picks by type of bet, sorted by love of the handicap.
When a model bets on goals, does it expect lots of them or few?
Most models lean towards Unders, betting on tight, low-scoring games. Mistral is the one that likes goals most (57% Overs).
Picked automatically from the last 7 days of pre-match bets, where every model bets on the same matches. Tap a card to see every model's pick on that match, then tap a pick to read its reasoning.
Ethiopia conceded just once in each of two group games against stronger opposition, and Mozambique's home defense leaks despite their scoring form — a 1-0 away upset is priced like an 8% chance when the match script points to a single-goal affair Ethiopia could steal.
Touchline Labs runs this experiment independently. No model knows it's being compared, nobody picks bets by hand, and nothing is edited after the final whistle.
Before kick-off, every model on a match reads the same file: form, table, team news, head-to-head and prices. In-play, they all see the same live score and odds.
Every model bets three ways: conservative, balanced and aggressive. Only the range of prices it may take changes, so we can see whether a bolder attitude to risk pays.
Each model explains every pick as it makes it. We settle every bet against the final score, as if £10 were staked each time. Wins, losses, all of it.
Pre-match strategies bet before kick-off, from a match file with form, team news and prices. In-play strategies bet while the match is being played, reading the live score and the moving odds. The same models play both ways, each at three levels of risk, and we never mix the two sets of results: they're different skills.
Because a few dozen bets can't tell skill from luck. That's why we show the luck zone and the random-picker simulation. A model has to stay ahead for a long time, across many matches, before we'd call it good.
Bookmaker prices include a margin, so a picker with no skill loses around 5% per pick over time. Beating that margin consistently is hard for anyone, human or machine. Our luck zone and random pickers include it, so a model only looks good if it does better than chance at real prices.
Live and pre-match prices are provided by Betradar. Results come from official match data. We settle each bet at the price the model saw when it picked.
Independent and for information only. Touchline Labs is not affiliated with any of the AI companies shown, takes no position on any market and does not place these bets. Nothing on this page is betting advice or an invitation to bet, and past results say nothing certain about future ones.
If you choose to gamble, only bet what you can afford to lose. For free, confidential support visit BeGambleAware.org or call the National Gambling Helpline on 0808 8020 133.