The AI Race on Polymarket: Who Wins, Priced in Real Time

The market map: how the AI race gets priced

Four families dominate the category. Best-model markets ask which lab tops a named benchmark leaderboard by a date - they are the category's flagship and resolve on a specific, named source, which matters more than casual traders realize. Release-date markets price whether a model ships by a deadline - binary, news-driven, and prone to violent repricing on a single tweet. Capability markets price specific achievements: passing a benchmark threshold, winning a competition, a model doing X by Y. Company-milestone markets cover valuations, revenue thresholds, IPOs and leadership changes - they trade more like equities-adjacent events than like model races.

Prices move on paper releases, leaked evals, conference keynotes and researcher departures - often hours before tech media writes the story. That speed is the whole point: the order book is where the field's collective read becomes visible first.

Resolution: the leaderboard IS the contract

AI markets have the sharpest resolution-criteria risk of any category. "Best model" is not an opinion in these markets - it is a number on a named leaderboard at a named timestamp. Which leaderboard matters enormously: rankings differ across evaluation sites, a model can lead one and trail another, and style-control or category filters can flip outcomes. Before trading any best-model market, read three things in the rules: the exact source, the exact reading time, and what happens in a tie or if the source goes down.

Release-date markets carry definition risk instead: what counts as "released"? Public general availability usually counts; a waitlist, a research preview or an API-only drop may or may not, and the rules say which. Traders who lost money on these markets almost always lost it on the definition, not the prediction. Our resolution guide covers the dispute process when an outcome lands in the gray zone.

Why AI markets move differently

Three structural differences from other categories. First, information is radically asymmetric: a tiny fraction of participants read evals, follow researcher Twitter, or can judge a technical report in an evening - and they consistently beat the price set by everyone else. Second, news is discontinuous: there is no polling average or game schedule smoothing the flow; a surprise launch moves a market 30 points in an hour. Third, narrative momentum overshoots: after a big release, the shipping lab's odds in every adjacent market get bid up beyond what the release logically implies, and the fade of that enthusiasm has been one of the category's most repeatable trades.

The practical consequence: limit orders matter more here than anywhere else. Fast markets punish market orders - the order book guide explains the mechanics.

The contenders: how the market frames the race

The live board below updates continuously, so this section describes structure. The market typically prices a two- or three-lab leading pack that rotates with release cycles, a fast-follower tier, and an open-source wildcard whose releases reprice everything because they reset the cost floor. Two patterns persist across cycles. Release cadence is priced, surprises are not: the market learns each lab's shipping rhythm and prices the next release in - what moves odds is shipping early, late, or better than the eval-leak whisper number. Benchmark leadership is sticky in price, fragile in fact: incumbency on a leaderboard earns a premium that one strong competitor release erases overnight - which is exactly the asymmetry sellers of incumbent-lab certainty have profited from.

Where the edges are

This category has the widest skill gap on Polymarket. Real edges, in rough order of accessibility: reading the actual evals - most participants trade headlines; the eval tables tell a different story often enough to pay for the reading time. Definition arbitrage - when a market's resolution criteria diverge from the headline framing people trade on, the mispricing is mechanical. Cross-market consistency - the same lab's odds across best-model, release-date and milestone markets occasionally imply contradictory worlds; the tighter market is usually right and the looser one pays. Conference calendars - odds drift predictably into major AI conferences and dev days, then resolve the drift in hours; the pattern repeats every cycle.

Size positions like the category is volatile, because it is - the position sizing guide applies double here.

AI markets as a forecasting instrument

Beyond trading, this category has become genuine research infrastructure: when labs, journalists and policy analysts want a base rate for "will AGI-adjacent capability X arrive by year Y", prediction-market prices are increasingly the first citation. The prices are not oracles - they inherit every bias of their participant pool, which skews young, online and optimistic about capability timelines. But they update faster than expert surveys, they are accountable in a way pundits are not, and their calibration track record on resolved AI questions has been respectable. Reading them well - as a probability with an error bar, not a verdict - is a skill worth having entirely apart from any position. Start with the probability thinking guide.