How accurate are the “best AI model” markets? A day out, Polymarket’s favourite won all 20 months; a month out, 16
Polymarket has asked every month since early 2025 which company will have the top model on the Arena text leaderboard, and $226 million has traded on the 20 monthly markets that have settled. We checked the favourite a month, two weeks, a week and a day before each check. A day out it never lost. A month out it lost four times, twice to companies priced at 3% or less.
In this article
Polymarket’s monthly “best AI model” markets have been very reliable close to the deadline and much less so a month out. We looked at all 20 monthly markets on which company has the top model on the Arena text leaderboard (formerly LMArena) that settled between February 2025 and October 2026, and compared the favourite at four points before each check with the company that won.
A day before the check, the favourite won 20 out of 20 times, at an average price of about 98%. A month before, it won 16 of 20, at an average price of 74%. Two of the four misses were companies priced at 3% or less a month out.
The short version
Source: Polymarket public market data and price history; PredictDesks analysis
- Big markets. The 20 markets traded about $226 million. The largest was the end-of-2025 market at $36.3 million, followed by January 2026 at $29.0 million.
- Three eras. xAI won the first market, at the end of February 2025. Google then won every monthly market from March 2025 to January 2026. Anthropic won eight in a row from February to September 2026. Google won again in October 2026.
- Close to the deadline, the favourite always won. A week out it won 19 of 20; a day out, 20 of 20.
- A month out, surprises happened. Four favourites lost, and in two of those months the winner was priced at 3% or less.
How the markets work
Each month Polymarket asks which company will own the model ranked first on the Arena text leaderboard at a set time on the last day of the month, usually noon ET. The rules use the overall text ranking with style control off, break ties on rank by Arena score, and in recent months exclude models marked “AutoEval.” The question is about the company, so any model from Google counts for Google.
The same question runs for the end of 2026, where Google is now the favourite at about 70% (live prices).
How we measured it
For each market we took Polymarket’s price history for every company with trading and read the price 30 days, 14 days, 7 days and 1 day before the end date. We scaled each set of prices to add up to 100% and called the company with the highest price the favourite. We then compared it with the company that won.
The sample is small: 20 markets, and consecutive months are not independent, because a model that leads at the end of one month often leads at the end of the next. We could not find a June 2025 market, so the series has one gap. All 20 markets were already trading 30 days before their check, so every market has a price at each of the four points. Volumes are Polymarket’s totals for each market, including trading after the result was clear. Treat the percentages as a description of what happened, not a precise forecast of what will.
How often the favourite won
Source: Polymarket price history; PredictDesks analysis
Show the dataHide the data
| Average price of the favourite | How often it won | |
|---|---|---|
| 30 days | 74.0% | 80.0% |
| 14 days | 87.6% | 90.0% |
| 7 days | 91.7% | 95.0% |
| 1 day | 97.8% | 100.0% |
| Time before the check | Favourite won | Average price of the favourite |
|---|---|---|
| 30 days | 16 of 20 (80%) | 74.0% |
| 14 days | 18 of 20 (90%) | 87.6% |
| 7 days | 19 of 20 (95%) | 91.7% |
| 1 day | 20 of 20 (100%) | 97.8% |
At every horizon the favourite won a little more often than its average price implied. With 20 markets that gap is within normal luck, but it points one way: on balance, these markets have not been overconfident about the leader.
The upsets
Source: Polymarket price history; PredictDesks analysis
Show the dataHide the data
| Value | |
|---|---|
| Feb 2025: xAI | 2.4% |
| Feb 2026: Anthropic | 3.0% |
| Mar 2025: Google | 10.7% |
| Aug 2025: Google | 46.0% |
Four months broke the pattern:
- February 2025. A month out, Google was priced at about 70% and xAI at about 2%. xAI won. A week before the check xAI had already become the favourite at 83%.
- March 2025. xAI was the favourite at 48% a month out and 91% a week out. Google won from about 11% a month out, the only time in our sample that the favourite a week before the check lost. A day out Google was at 98%.
- August 2025. OpenAI led narrowly a month out at 51% against Google’s 46%. Google won, and was the favourite at 93% two weeks out.
- February 2026. Google was the favourite at 74% a month out and Anthropic at about 3%. Anthropic won, and began a run of eight straight months at the top. Two weeks out Anthropic was already 73%.
The timing varied. In August 2025 and February 2026 the eventual winner was already the favourite two weeks out. In February 2025 it took until a week before the check, and in March 2025 the leader changed only in the final week. Once a new company reached the top of the board, the market followed it quickly.
Long runs at the top
The markets also show how long a lead lasts. Google won all 10 monthly markets in our sample from March 2025 to January 2026, and was priced at 84% or more a month out in the last five of them. Anthropic’s run from February to September 2026 was similar, at 80% to 93% a month out from June onwards.
Long leads make the favourite a month out a good bet most of the time. Late in each run, the favourite traded at 90¢ or more for most of the month, which leaves little to gain from backing it.
How well the prices were calibrated
Source: Polymarket price history; PredictDesks analysis
Show the dataHide the data
| Value | |
|---|---|
| Priced 80% or more | 100.0% |
| Priced 50% to 80% | 62.5% |
| Priced 20% to 50% | 20.0% |
| Priced 5% to 20% | 7.1% |
| Priced under 5% | 1.0% |
Looking at every company in every market, not just the favourite, the prices a month out were close to right across the scale. Companies priced at 80% or more won all 10 times. Companies priced between 50% and 80% won 5 of 8, close to their average price of 64%. At the bottom, the 195 company prices under 5% produced 2 winners, against about 1.6 implied by their prices.
A week out the picture was sharper. The 18 company prices of 80% or more won 17 times, against an average price of 94%; the 218 prices under 5% produced a single winner, the March 2025 upset.
The markets carry many companies, about 12 per market on average, but nearly all of the money and nearly all of the wins went to three. Google’s winning months account for about $111 million of the $226 million traded and Anthropic’s for about $110 million.
When the money comes in
Source: Polymarket public market data
Show the dataHide the data
| Value | |
|---|---|
| Feb 25 | $5.6m |
| Apr 25 | $3.8m |
| Jul 25 | $6.1m |
| Sep 25 | $2.9m |
| Nov 25 | $7.5m |
| Dec 25 | $36.3m |
| Jan 26 | $29.0m |
| Feb 26 | $22.5m |
| Mar 26 | $16.2m |
| Apr 26 | $21.1m |
| Jun 26 | $22.3m |
| Aug 26 | $3.3m |
| Oct 26 | $4.2m |
Trading peaked around the turn of the year. The end-of-2025 market drew $36.3 million, January 2026 $29.0 million and February 2026, the month of Anthropic’s upset, $22.5 million. By late summer the monthly markets had shrunk to $3 million to $4 million, as Anthropic’s lead looked secure. October 2026, when Google took the top spot back, drew $4.2 million.
What this means for the year-end market
The market on the end of 2026 has Google at 70.5% with about seven weeks to go, and November’s market has Google at 75.5%. Our data suggests three things:
- A favourite this far out wins most of the time but not always. At 30 days, favourites at an average price of 74% won 80% of the time; at seven weeks the uncertainty is greater.
- Upsets can come fast. In February 2026 Anthropic went from about 3% a month out to the favourite at 73% two weeks out.
- The last week is safe ground. A week out, only one favourite lost in 20 months; a day out, none did.
What the study does not cover
- Second and third place. Polymarket also runs markets on which company has the second-best and third-best model each month. They are much smaller, from about $86,000 to $2.6 million for the second-best markets, and we have not tested them here.
- Coding and maths. Since April 2026 Polymarket has listed separate markets on the best coding and maths models, on different leaderboards.
- Kalshi. Kalshi runs a similar question by product name (Gemini, Claude, ChatGPT and others), with monthly, weekly and year-end versions. We used Polymarket because its price history covers all 20 months.
- The leaderboard itself. We test the markets, not the leaderboard. Whether a crowd-voted ranking is the right measure of the “best” model is a separate question; the markets simply settle on it.
How to use these results
If you trade the year-end market or the monthly ones, three practical points follow from the data:
- A favourite above 80% a month out has not lost. All ten such prices won. That does not make it certain, but it suggests the strongest leads are rarely overturned within a month.
- Prices in the middle are where the risk is. Favourites priced between 50% and 80% a month out won five times in eight. If you hold one, a new release can change everything within days.
- Cheap long shots have occasionally paid. Two companies priced under 3% a month out went on to win. Across 195 such prices, two winners is roughly what the prices implied, so there is no free lunch, but there is a real chance of an upset.
FAQ
Which company has won the most monthly markets? Google, with 11 of the 20 that have settled, followed by Anthropic with 8 and xAI with 1.
How far ahead can the markets call the winner? A week before the check, the favourite won 19 of 20 times; a day before, all 20.
What was the biggest upset? By price a month out, xAI in February 2025 at about 2.4%, and Anthropic in February 2026 at about 3%.
Which leaderboard do the markets use? The Arena text leaderboard, overall ranking, with style control off. The leaderboard was known as LMArena (Chatbot Arena) for most of the period we studied; the newer markets refer to arena.ai.
Is the November 2026 market different? No. It uses the same rules, with the check at noon ET on 30 November. Google is the favourite at 75.5%. On the evidence of the 20 earlier months, a favourite at that price two to three weeks out has usually, but not always, gone on to win.
Data
All 20 markets, with the favourite and its price at each point, the winner and the volume, are free to download: best-ai-model-markets-2025-2026.csv. For accuracy across other kinds of market, see our prediction market accuracy study.
Markets settled from February 2025 to October 2026. Prices from Polymarket’s public price history at 12-hour resolution. Nothing here is financial advice.
