When Every AI Agrees on a Pick, Trust It Less. We Have the Data.

Last Monday there were seven MLB games where all twelve AI models on our leaderboard picked the same team. Every single model, same side, seven times.

Those unanimous picks went 2-5. The loudest one was the Braves at the Mets, where twelve models took Atlanta and the Mets won 14-3.

If your instinct says a full sweep of the board should be the strongest possible signal, you're in good company, because that was our instinct too. We've now graded a month of these and the data says the instinct is backwards, and a month of grading unanimous picks explains exactly why.

The numbers first

From June 30 through July 25 there were 180 graded games where every reporting model (minimum four, usually nine to twelve) picked the same moneyline side. Betting every one of those unanimous calls returned +7.4% per pick. Decent, right?

Here's the problem. Betting just TWO models, Claude Opus 4.8 and Claude Sonnet 5, whenever they agreed with each other returned +12.6% over the same period. Adding ten more agreeing models to that pair didn't strengthen the signal. It cut it nearly in half.

The totals version is worse. When the whole field projects the same side of the over/under, the unanimous unders have already flipped negative in recent weeks, and the unanimous overs are 26-26 lifetime, a literal coin flip. Full agreement of a dozen AI models on an over carries literally zero information against the price.

Why more agreement means less edge

The mechanism is almost embarrassing once you see it. Of those 180 unanimous games, 167 were on the favorite. Unanimity doesn't measure how much the field knows. It measures how OBVIOUS the game is. When twelve models all land on the same team, they're usually all reading the same lopsided matchup that the betting line already priced. You're not getting twelve independent opinions, you're getting the same opinion twelve times, at a price that already knows it.

The market is the thirteenth forecaster in the room, and it's the sharpest one. A pick only pays when it disagrees with the price, and obvious games are exactly where the price is hardest to beat.

The disagreement flip

Now run it the other way. Take the games where the two Claudes agreed with each other but at least one OTHER model took the opposite side. Contested picks, 55 of them in our window. Those returned +18.8%, the best consensus cut we've measured, at prices near even money.

Same two models. Same agreement. The only difference is whether the rest of the field fought them, and the fight is where the money was. Their hit rate barely moved between contested and uncontested games. The PRICE moved, because contested games are the ones the market hasn't already decided.

There's one more wrinkle we didn't expect: the Claudes' own confidence works backwards too. Their hesitant agreements, the ones where both models sat at 55% or below, outperformed their confident ones by a wide margin. We liked that cut enough that it's now its own tracked strategy on the Pro board, record and equity curve public like everything else.

What we do with this

Practically, three things changed on our end. The strategies we track are built on narrow signals, two specific models or one specific model's specific flaw, not on field-wide agreement. The daily unanimous count still gets published because it's honest data, but nobody should read it as conviction. And when we compare any two pickers head to head, we grade the games where they DISAGREED, because that's where information lives. You can run that comparison yourself for any two models or experts on our head-to-head pages.

Monday's 2-5 was no anomaly, just a small sample of a pattern we'd already measured across 180 games: consensus is a proxy for obviousness, and obviousness is already in the price. Twelve models agreeing tells you the game looks easy. It has never told anyone who wins.

The full methodology is on the accuracy page, the essay-length version of whether any of this beats the market is at Can AI Beat Vegas?, and every graded prediction is in the open dataset if you want to check the unanimity math yourself.