LMArena Odds 2026: Gemini Pro Priced at 20%

LMArena odds are priced as score ladders, and a ladder is supposed to be readable at a glance: the harder the bar, the lower the price. The Gemini Pro board on Polymarket has $341,765.90 riding on six such bars and it does not obey that rule, because four of its six score bands currently price at negative probability. We read all three live Arena debut boards from the Gamma API at 21:10 UTC on 28 October 2026, pulled thirty days of CLOB price history for every leg, and read each leg's resolution text against Arena's own leaderboard changelog. The short version is that the Gemini board stopped being a score forecast on 1 October 2026, and nobody has repriced it as the thing it now is.

That matters because these boards look like the most legible markets in AI: a score threshold is objective and the resolution source is a public leaderboard. Three of the four findings below are reasons to trust the prices less than that framing suggests.

1. What an Arena debut market actually asks.

2. The Gemini Pro board collapsed 52 points on 1 October.

3. Four of the six bands now price at negative probability.

4. Why the ladder flattened: the market needs a model called Pro.

5. The 1470 leg is the tell.

6. A second board, a different benchmark, the same 20%.

7. GPT Astra's ladder is the one that behaves.

8. Grok's spreads turn its prices into decorations.

9. Six checks before you trade an Arena debut market.

An Arena debut market is not a bet on which model is best. It is a bet on the number that appears in one column of one leaderboard tab on one particular afternoon. Each board is a set of separate Yes/No contracts at rising score strikes, so the Gemini board runs 1470+, 1480+, 1490+, 1495+, 1500+ and 1510+, and you trade whichever bar you have a view on.

The resolution wording is tighter than most traders assume. The score comes from the Score column under the Text Arena Overall tab with style control off, read at 12:00 PM ET on the calendar date following the day the qualifying model first appears. Not the day it is released, and not the day it settles. On the GPT Astra and Grok boards a model flagged AutoEval does not count as added at all, and if several models arrive on the same date the highest scorer is used.

Two further clauses decide more of these markets than the strikes do. A model added on the calendar date after the first qualifying model is ignored, so only one debut per board ever counts. And every board has a backstop date: if no qualifying model appears by 31 Dec 2026 for the Gemini board, or 30 June 2027 for the Astra and Grok boards, every leg resolves No.

That backstop is the clause to read first, because it means a leg can lose without the model ever missing its score. Most of the time that is a minor tail risk. On one of these three boards it has become the entire price.

Thirty days of CLOB history make the break easy to date. On 29 September 2026 the board was a textbook ladder, descending in order: 1480+ at 94.70%, 1490+ at 91.90%, 1495+ at 84.00%, 1500+ at 68.00% and 1510+ at 45.10%. Top to bottom it spanned 49.6 points, which is what a real score forecast looks like. On 30 September it held, closing at 94.90%, 91.90%, 83.50%, 73.00% and 38.40%, spanning 56.5 points.

On 1 October it fell apart. The 1480+ leg closed at 42.80%, down 52.1 points in a day. The 1490+ leg went to 38.70%, the 1495+ leg to 45.50%, the 1500+ leg to 41.50% and the 1510+ leg to 32.20%. The spread from easiest to hardest bar compressed from 56.5 points to 13.3 points, and the ordering broke the same day.

It never recovered. Through 28 October the six legs ground down into a single cluster, and today they sit between 17.80% and 22.00%, a span of 4.2 points across forty points of score. The ladder is now flat, and a flat ladder is an inverted one the moment bid-ask noise exceeds the gaps between its rungs.

Counting properly: across the thirty days we have history for, the board was monotonic on six of them, 29 and 30 September plus 4, 5, 7 and 8 October. Since 1 October it has been non-monotonic on 24 of 28 days. Today's reading of four simultaneous violations is the worst in the series.

A cumulative ladder carries more information than its individual prices, because the gap between two adjacent strikes is the probability of landing between them. Subtract 1480+ from 1470+ and you get the implied chance of a debut in [1470,1480); do that down the board and you should get a distribution over the score.

On the Gemini board you do not. Today's mid prices are 18.50% at 1470+, 20.35% at 1480+, 20.80% at 1490+, 21.00% at 1495+, 22.00% at 1500+ and 17.80% at 1510+. The implied bands come out at negative 1.85% for [1470,1480), negative 0.45% for [1480,1490), negative 0.20% for [1490,1495) and negative 1.00% for [1495,1500). Only [1500,1510) at 4.20% and 1510+ at 17.80% are positive, alongside 81.50% for a score below 1470.

Negative probabilities are not a modelling choice; they are the arithmetic telling you the prices cannot all be right. The four violations are 1470+ priced 1.85 points below 1480+, 1480+ below 1490+ by 0.45, 1490+ below 1495+ by 0.20, and 1495+ below 1500+ by 1.00. Cumulatively the board says a 1500 debut is 3.50 points likelier than a 1470 debut, which cannot be true, because every score at or above 1500 is also at or above 1470.

It is executable too, which rules out rounding. Buying 1470+ Yes at its 0.190 ask and 1500+ No at its 0.800 ask costs 0.990 for a pair that returns at least 1.00 in every state of the world, and 2.00 if the debut lands between the two bars. Two other pairs price the same way. The catch is size: the 1470+ leg carries $1,105.26 of liquidity, so the locked cent is not a trade. Its value is diagnostic. Nobody is enforcing coherence on this board, so no individual price on it should be read as a considered forecast.

The explanation is in the first paragraph of the resolution text, and it is a label rather than a number. The market resolves on the next Google Gemini Pro model added to the leaderboard. Any Gemini newly added and labeled Pro may qualify, and the wording offers gemini-2.5-pro, gemini-3-pro and gemini-3.1-pro-preview as its examples. Models labeled only Flash, Flash-Lite or another non-Pro variant are explicitly excluded.

On 30 September 2026 Google announced Gemini 4 Argon as its most advanced model, and Arena's changelog records gemini-4-argon-high being added to the Text Arena that same day. Several outlets describe Argon as the first Gemini to carry a codename in place of the Pro and Flash tier labels, though Google has not said what the rest of the Gemini 4 line will be called, so treat the tier story as reported rather than confirmed. Engadget adds the detail that matters most here: a Gemini 3.5 Pro was planned for earlier in 2026 and shelved when Google moved its focus to Gemini 4.

So the frontier Gemini arrived on the resolution source and did not trigger the market, because it is not called Pro. The board fell 52 points the next day. Arena's changelog lists every Text Arena addition through 22 October, and the only Gemini entries are Gemini 3.8 Flash (High) on 2 September and gemini-4-argon-high on 30 September; the 6 October gemini-nano-banana-2.1 entry is an image model. No Pro-labeled Gemini has been added at all.

That turns six score bets into six noisy copies of one binary question: will Google ship something it still calls Pro, and will Arena list it, inside the nine weeks to 31 Dec 2026? The strikes stop mattering, every leg converges on the same probability of roughly a fifth, the gaps between rungs shrink below the spreads, and the ordering starts crossing at random. The market wording is a fossil of a naming scheme its subject has moved on from.

If you want one number that proves the board is pricing the label and not the score, it is the easiest bar on it. The 1470+ leg trades at 18.50%, and 1470 is not a demanding threshold for a Gemini flagship. Reporting on the Gemini 3.1 Pro launch in February 2026 put Gemini 3 Pro at 1470 on the text leaderboard, and Arena's own launch post said Gemini 3.1 Pro arrived tied for first in Text at 1500.

The frontier has moved up since. Third-party trackers of the 20 October 2026 text leaderboard put Gemini 4 Argon (High) at the top, though they disagree on the figure, quoting between 1525 and 1532; the same snapshots put Claude Opus 5.5 (High) at 1516 and Claude Fable 5.1 (Max) at 1512. Treat the exact level as approximate, because arena.ai returns 403 to a fetcher and we could not read the live board directly.

The direction is not in doubt, though. A new Gemini flagship clearing 1470 is close to a formality when the previous Pro generation already scored 1470 and 1500, and the current top of the board is around 1530. Conditional on a Pro-labeled Gemini actually being added, 1470+ should price in the nineties.

It prices at 18.50%. Almost all of that discount is the probability that no qualifying model appears at all, which means the board is quoting roughly an 81.50% chance that Google ships nothing called Pro onto the leaderboard before the year ends. That is a defensible view. It is simply not a view about scores, and the six strikes dressed around it are decoration.

If the label really is doing the work, the same signature should appear on any other market gated on the next Gemini Pro model, whatever it measures. It does, on a board that has nothing to do with Arena scores.

Polymarket's Next Google Gemini Pro Model: Humanity's Last Exam Debut board asks for a score on an entirely different benchmark, and it carries $118,636.51. Its four legs price at 20.75% for 40%+, 21.50% for 45%+, 20.00% for 50%+ and 13.50% for 55%+. The same plateau around a fifth, and the same defect: 45%+ trades 0.75 points above the strictly easier 40%+.

Two independent benchmarks, two unrelated score scales, one shared clause and one shared deadline of 31 Dec 2026, both plateaued near 20% and both crossing themselves. That is strong evidence the common factor is the Pro-label contingency rather than anything about Arena, Elo or Humanity's Last Exam.

Taken together the two boards have $460,402.41 riding on a naming decision at Google. Neither is mispriced in the sense of being obviously too high or too low; both are miscategorised. They are listed and read as benchmark markets, and they are trading as a single binary about product branding.

The Gemini board is not evidence that Arena debut markets are broken as a format, and the cleanest proof is the board sitting next to it. The GPT Astra 6.1+ Text Arena board holds $50,750.16 across five strikes and descends in order, exactly as a cumulative ladder should: 81.50% at 1430+, 63.00% at 1440+, 53.50% at 1450+, 43.00% at 1460+ and 18.50% at 1470+.

Because it is monotonic, the subtraction works and you get a genuine distribution. The implied bands are 18.50% below 1430, 18.50% for [1430,1440), 9.50% for [1440,1450), 10.50% for [1450,1460), 24.50% for [1460,1470) and 18.50% at or above 1470. They sum to exactly 100%, and the modal outcome is a debut in [1460,1470) at 24.50%.

The spreads support that reading rather than undermining it. Four of the five legs quote 1 or 2 points wide, with $5,979.25 of liquidity behind the board, so the mid prices are close to what you could actually transact. The difference from the Gemini board is not volume, which is six times smaller here; it is that the question still has a live score dimension. OpenAI's naming convention is what the market tracks, and GPT Astra 6.1+ is a version rule rather than a tier label, so the model that resolves it does not have to be called anything in particular.

That is the useful contrast. Read side by side, the Astra board tells you what traders expect the next OpenAI flagship to score, while the Gemini board tells you what they make of Google's naming. Same format, same resolution source, different questions.

The third board shows a different failure, and it is the one most likely to catch a reader who trusts a screenshot. The Next Grok Model (4.8+) Text Arena board quotes mid prices that look perfectly orderly: 45.00% at 1440+, 39.00% at 1450+, 37.50% at 1460+ and 17.50% at 1470+. Monotonic, plausible, and almost entirely meaningless.

Look at the books instead of the mids. The 1440+ leg is 0.300 bid against 0.600 offered, a 30-point spread. The 1450+ leg is 0.140 against 0.640, a 50-point spread. The 1460+ leg is 0.120 against 0.630, 51 points wide. The 1470+ leg is 0.050 against 0.300. The whole board holds $592.99 of liquidity and $6,804.64 of lifetime volume, with one leg having traded $639.66 in total.

A 45.00% mid price derived from a 30 to 51 point spread is not a forecast, it is the midpoint of an absence. Anything between roughly 30% and 60% is consistent with that book on the 1440+ leg, which is wide enough to swallow the entire difference between the strikes. The ladder looks coherent here only because nothing has pushed on it hard enough to break it.

This is the ordinary condition of small AI boards, and it is worth internalising before quoting any of these numbers. A clean-looking ladder on a thin book tells you less than an ugly one on a deep book.

First, read the qualifying clause before the strike. Ask what makes a model count, and whether it is a version rule like GPT Astra 6.1+ or a tier label like Gemini Pro. Label rules can be retired by a product decision; version rules usually cannot.

Second, check the backstop date and ask what share of the price it explains. If the easiest strike on the board trades far below where the subject's own previous generation already scored, you are looking at an appearance market wearing a score market's clothes.

Third, subtract adjacent strikes and look for a negative band. It takes a minute and it is the fastest test of whether anyone is arbitraging the board. Any negative band means at least two prices on it are wrong, and you cannot tell which from the board alone.

Fourth, price from the bid and the ask, never the mid. On the Grok board the mid is the midpoint of a 51-point gap. Fifth, remember the measurement is a snapshot: resolution reads the score at 12:00 PM ET the day after a model first appears, while Arena scores drift as votes accumulate. Gemini 3.1 Pro was reported at 1500 on launch and at figures from 1487 to 1505 in later snapshots, with stated intervals of several points, so strikes set 5 points apart at 1490 and 1495 are closer together than the noise in the number they resolve on.

Sixth, treat a leaderboard rename as a resolution risk. This board's own text cites lmarena.ai in one paragraph and arena.ai in another, a residue of the January 2026 rebrand, and lmarena.ai now redirects to arena.ai. None of that has broken a settlement yet, but it is the same class of drift that turned the Pro label into a dead clause.