Free · For fun

Bubble AI

NFL matchups, projected scores and market lines in one place. Explore the numbers behind each game and make your own call.

For fun, not betting advice. Bubble AI is a free computer model built for entertainment. We don't take bets or link to sportsbooks. Read our model testing and limitations. If gambling stops being fun, call or text 1-800-GAMBLER. More in our Terms.

Think you can beat Bubble AI?

Take $1,000 in pretend money to this week's real games and see how high you can run it. Free, no real money in, no prize out: bragging rights only.

Play the $1,000 Challenge

This week

Loading this week's games…

Make your own call

Think the model's wrong? Pick every game straight up in our free weekly contest.

Play NFL Pick'em

Bubble AI Power ratings

How it works

Every team gets an offense and a defense rating from this season's final scores, adjusted for who they played. Last season counts a little early on and fades out.
For each game the model projects both scores, adds home field, and docks a team starting a backup quarterback.
It compares its line with the market's and shows the side it leans to. The bars show how far apart they are, not how likely the lean is to hit.
A very big gap usually means the market knows something the model can't see, like an injury.
The moneyline chance shown first is the market's own, with the house cut removed. The model's estimate is shown beside it for comparison. The testing notes below explain the difference.
Player props: each player's projection is 60% his season average and 40% his last three games, nudged for the defense he faces, shown next to the posted line. Prop picks have not been tested against real lines (no free history exists), so treat them as a talking point.
Biggest Disagreements lists the games on one day where our line and the market's are furthest apart, leaving out gaps so big the line likely knows something.

Model testing and limitations

Our backtest covered graded regular-season games from 2018 through Week 7 of 2026: 2,176 spread picks hit 49.7%, and 2,214 total picks hit 48.2%. Both were below the 52.4% break-even rate at -110; bigger gaps performed worse. This test found no winning edge, so spread and total leans are not presented as calibrated cover probabilities. The market's moneyline chance was better calibrated than the model's in this test.

Historical lines were approximately closing lines, while this page compares current lines. The backtest used prior-game data; its quarterback adjustment used the actual starter, which may not always have been known before kickoff. These results describe the existing score-based model, not a future replacement.

Team ratings from final scores in nflverse data (CC BY 4.0), updated weekly. Market lines, player prop lines and schedule from ESPN's public feeds, shown for comparison only. Player projections from nflverse weekly stats. Not affiliated with any sportsbook.