Which AI predicts esports best? Models ranked by real bets
We compare AI models on Dota 2 and CS2 bets recorded before the series starts. For each model we count how many bets won, the average odds, the ROI and profit to a one-unit stake. Below are the results by game and by market, plus the bets on upcoming matches.
There are 16 AI models in the esports ranking right now. Since 24 September, 217 bets have been settled across 15 matches: 54% won, for an overall ROI of +2.7%. For comparison, backing the favourite in the same matches returned an ROI of -12.8%. MiniMax leads on profit (+4.61 units), but the sample is still small and the standings can change after every tournament day.
Ranking by profit
| # | AI model | Bets | Pending | Won | Avg odds | ROI | CLV | Profit |
|---|---|---|---|---|---|---|---|---|
| 01 | MiniMax | 19 | 8 | 8 of 11 | 2.02 | +41.9% | — | +4.61 |
| 02 | Mistral | 21 | 7 | 10 of 14 | 1.93 | +31.5% | — | +4.41 |
| 03 | Kimi | 24 | 9 | 9 of 15 | 1.97 | +19.9% | — | +2.99 |
| 04 |  GLM | 24 | 9 | 9 of 15 | 1.90 | +18.2% | — | +2.73 |
| 05 | Claude | 24 | 9 | 9 of 15 | 1.93 | +13.2% | — | +1.98 |
| 06 | DeepSeek Flash | 22 | 9 | 7 of 13 | 2.20 | +14.4% | — | +1.87 |
| 07 | Nemotron Super | 20 | 9 | 7 of 11 | 2.00 | +17.0% | — | +1.87 |
| 08 | DeepSeek | 24 | 9 | 8 of 15 | 1.98 | +2.4% | — | +0.36 |
| 09 | Qwen | 24 | 9 | 8 of 15 | 2.00 | +1.1% | — | +0.17 |
| 10 |  Gemini | 24 | 9 | 8 of 15 | 1.92 | -0.2% | — | -0.03 |
| 11 |  Sonar | 23 | 8 | 8 of 15 | 1.75 | -2.8% | — | -0.42 |
| 12 | ChatGPT | 24 | 9 | 7 of 15 | 1.93 | -12.5% | — | -1.87 |
| 13 |  Dots3-Note | 18 | 8 | 4 of 10 | 2.06 | -19.7% | — | -1.97 |
| 14 |  Grok | 24 | 9 | 7 of 15 | 1.92 | -15.6% | — | -2.34 |
| 15 | Xiaomi MiMo | 24 | 9 | 6 of 15 | 1.91 | -24.1% | — | -3.61 |
| 16 | Nemotron Ultra | 17 | 9 | 2 of 8 | 1.72 | -59.9% | — | -4.79 |
| All AI models together | 356 | 139 | 117 of 217 | 1.94 | +2.7% | +0.2% | +5.96 | |
| AI consensusthe most common bet | 15 | 8 of 15 | 1.92 | -0.2% | — | -0.03 | ||
| The favouritethe side with the lower price to win the series | 15 | 10 of 15 | 1.46 | -12.8% | — | -1.92 |
Ranked by profit in units to a 1-unit stake; won excludes voids; ROI is profit per unit staked; pending bets are on series not yet played.
CLV shows how much higher our odds were than the last price before the series; that price was available for 215 of 217 settled bets, and a row displays it once a model reaches 20 bets with a closing price.
The comparison rows are calculated on the same matches; the favourite's price is the last one before the series starts.
There are few series so far: the standings can change after every tournament day.
AI model results by game
Each cell shows won out of settled; from 10 settled bets the colour indicates profit: green for profit, red for a loss.
A game name leads to its upcoming matches and predictions.
What AI models bet on and what wins
| Market | Share of bets | Won | Avg odds | ROI |
|---|---|---|---|---|
| Map handicap | 53% | 88 of 143 | 1.94 | +16.3% |
| Series result | 37% | 15 of 52 | 1.90 | -45.4% |
| Total maps | 10% | 14 of 22 | 2.05 | +28.2% |
Share of all AI bets, including series not yet played; won and ROI are based on settled bets.
How we compare AI models
One match card. For every series the models get the same data: current rosters and how long they have played together, form by series and maps, the hero pool (Dota 2) or the map pool (CS2), rankings, the head-to-head record and the odds. They do not see each other's answers. Sonar can also use web search, and this is marked separately.
One bet per model. Each AI model picks one available market: the series winner, a map handicap or total maps. If we have no odds for the chosen market, the bet does not count.
The same odds for every model. We settle at the price recorded when the bets are collected: the average bookmaker odds, or the prediction-market price where there are no bookmaker odds. The price the model named itself is not used, so every participant is judged by the same rules.
Ranked by profit. Rankings depend on the odds taken as well as how many bets won. We also show ROI, average odds and results in each game.
Esports has its own table. Football and esports bets are never combined: the markets and prices are different, so each sport is ranked separately. Dota 2 and CS2 share one table, with the results for each game shown below it.
Switching models does not reset the record. A participant is one developer's AI. When we move it to a different model, its bets keep counting towards the same total.
We do not publish the models' full answers. The match page shows each AI model's bet and the consensus, their most common bet. The models' detailed reasoning is not part of the ranking.
Questions about AI esports bets
Which AI predicts esports best?
MiniMax leads on profit across Dota 2 and CS2 right now: 8 of 11 won, returning +4.61 units to a 1-unit stake. It is too early to declare a clear winner: the models have between 8 and 15 settled bets each, and the rankings can change after every tournament day.
Can you bet on AI esports predictions?
Here is the record so far: all AI models combined returned an ROI of +2.7% across 217 bets; backing the favourite in the same matches returned -12.8%; the AI consensus (the most common bet) returned -0.2% across 15 matches. These are past results, not a promise: betting means risking money, and the sample is still small.
How do the AI models differ from each other?
In style — what they bet on and at what prices: 50% of Gemini's bets are on the series result; 81% of Mistral's bets are on a map handicap; DeepSeek Flash has the highest average odds (2.20) and Sonar the lowest (1.75).