LiveBench Odds 2026: Anthropic 5-0 on Polymarket

LiveBench odds are the quietest corner of the AI prediction-market board and the only one with a clean track record. Polymarket has settled five markets on the LiveBench leaderboard since August 2026, Anthropic won all five, and six more are open with Anthropic priced between 66.50% and 93.00%. That matters because almost every other AI ranking market disagrees with itself from one month to the next.

We read all six live boards from the Polymarket Gamma API at 16:10 UTC on 24 October 2026, pulled every leg's price history from the CLOB, recovered the five resolved markets through Gamma's resolved-event search, and checked the resolution source against three independent LiveBench trackers, because livebench.ai itself renders client-side and serves no readable leaderboard to a fetcher.

Four things fell out, and three are reasons to be more careful than the headline prices suggest. Eleven dead legs on every board make the naive probability sum nonsense. The favourite on the biggest board moved almost twenty points in a day. And the benchmark these markets name has had a version break the wording has only half absorbed.

The short version: the streak is real and the leaderboard gap behind it is real, but the price you can actually trade is set by a thin book with wide spreads, and two of the six boards resolve on a column that may no longer be on the page.

1. Anthropic has won all five LiveBench markets settled so far.

2. What the six live LiveBench odds boards price right now.

3. What a LiveBench market actually resolves on.

4. The LiveBench 2.0 break and the Mathematics column.

5. Eleven dead legs and a board that sums to 666.5%.

6. The favourite moved 19.65 points in a single day.

7. Spreads turn half of these prices into decorations.

8. Six checks before you trade LiveBench odds.

Start with the record, because no other page on this subject puts it in one place. Polymarket ran LiveBench boards for August and September 2026 and every one resolved to Anthropic.

August Overall resolved Anthropic on $30,117.51 of volume and August Coding on $51,410.48, both settling 1 September 2026. September Overall resolved on $68,018.36, September Mathematics on $56,104.09 and September Coding on $176,431.13, all settling 30 September. Five markets, five Anthropic wins, $382,081.58 of combined turnover.

September Coding is worth sitting with. At $176,431.13 it is the largest market the series has produced, bigger than the entire October trio put together ($146,464.31), and the losing side was not a rounding error: OpenAI's leg took $25,887.86 against $27,829.43 on the Anthropic leg that paid out. Real money took the other side of a 5-for-5 streak and lost.

That is the frame for everything below. A market that has resolved the same way five times is not one with an unknown winner; it is one where the question has quietly become what the streak is worth as a price. At 93.00% on October Coding you pay 93 cents to make 7, and the five previous settlements are already inside that number.

Add the live boards to the settled ones and the series has turned over $566,303.44 since August. That is small beside the arena-rank markets asking which company has the best model overall, which is exactly why these boards behave differently: fewer traders, tighter leaderboard evidence, and a resolution source you can read yourself.

Six boards are open, three for October and three for November, and they split into two very different groups. Prices and volumes were read from Gamma at 16:10 UTC on 24 October 2026.

October, checked 31 October. Overall: Anthropic 89.40%, OpenAI 0.75%, Moonshot 0.60%, Meta 0.45%, DeepSeek 0.45%, Google 0.20%, on $58,566.18 of volume and $35,207.39 of liquidity. Mathematics: Anthropic 89.00%, OpenAI 4.95%, Google 2.55%, on $55,680.19 and $30,917.07. Coding: Anthropic 93.00%, OpenAI 3.55%, Google 1.40%, Meta 1.00%, on $32,217.94 and $15,967.82.

November, checked 30 November. Overall: Anthropic 74.50%, Tencent 10.55%, Meituan 5.90%, Google 5.40%, OpenAI 5.40%, on $11,418.60 and $13,430.18. Mathematics: Anthropic 66.50%, OpenAI 17.50%, Google 5.15%, on $19,090.34 and $15,445.40. Coding: Anthropic 80.50%, Moonshot 10.75%, Google 5.50%, OpenAI 5.25%, Nvidia 5.25%, on $7,248.62 and $3,460.48.

Two patterns. The October trio holds $146,464.31 against $37,757.56 on the November trio, a ratio of 3.88 to 1, which is what a week to run looks like beside five. And the favourite's price decays hard with time: 89.40% to 74.50% on Overall, 89.00% to 66.50% on Mathematics, 93.00% to 80.50% on Coding. The market is not saying Anthropic weakens in November; it is pricing five extra weeks in which somebody could ship something.

Coding is the strongest board in both months and Mathematics the weakest, four points apart in October and fourteen in November. If you think the next frontier release lands in a reasoning-heavy category rather than a coding one, November Mathematics at 66.50% is where that view is cheapest to hold.

All six boards resolve the same way, and the wording rewards reading twice. The market resolves to the company owning the model with the highest score in the named column of the leaderboard at livebench.ai, when the board is checked at one stated moment, with the latest available LiveBench release selected and the category set to All, Coding or Mathematics.

The October boards are checked on 31 October 2026 at 12:00 PM ET; the November boards on 30 November at 12:00 PM ET. There is no averaging and no window. One read of one page decides all three boards in a month, which is why a release landing on 29 November matters far more to these markets than one that landed on 2 November and has already been scored.

The tiebreaks are where this gets unusual. If two models show exactly the same score, the one with the lower listed cost per successful task ranks ahead. If that still ties, company names are ordered alphabetically as listed in the market group, and the market's own example is that Google would rank ahead of SpaceXAI. That clause also tells you how the group names xAI, worth knowing before you hunt for an xAI leg that is not labelled that way.

One timing detail is inconsistent across the series and nobody has flagged it. The October boards stop trading at 16:00 UTC on 31 October, which is exactly 12:00 PM ET, so the book closes on the deciding snapshot. The November boards stop at 04:59 UTC on 1 December, which is 11:59 PM ET on 30 November, roughly twelve hours after their check. On the November boards the outcome is knowable, from the same public page the market names, for half a day while the book is still open.

Finally the source-failure clause: if the leaderboard is unavailable at check time the market stays open until it returns and resolves on the first check after, and if it goes permanently offline the market resolves to Other. That is the only path on which one of those eleven phantom legs pays.

Here is the part that should make a Mathematics position uncomfortable. LiveBench has had a version break, and the market wording sits on both sides of it.

LiveBench 1.0 was frozen at its final release on 25 June 2026. A third-party mirror of that release lists 71 model variants across 23 objective tasks in seven categories: reasoning, coding, agentic coding, mathematics, data analysis, language and instruction following. LiveBench 2.0, which livebench.ai serves today, describes itself as a contamination-free long-horizon benchmark that has models produce deliverables such as decks, research reports, bug fixes and multi-hour apps, then grades the output. Its categories are agentic, reasoning, coding, browser use, data analysis, research and design, and it carries a cost per successful task column.

Note which clause depends on which version. The tiebreak names cost per successful task, a 2.0 column, so the markets were written against the 2.0 board. But two boards resolve on a Mathematics column, and mathematics is a 1.0 category absent from the 2.0 list. Those two hold $74,770.53 between them: $55,680.19 on October Mathematics and $19,090.34 on November Mathematics.

Be clear about the strength of this evidence, because it is the kind of claim that gets repeated carelessly. We could not read the live leaderboard directly, so the 2.0 category list here comes from the site's own indexed description and the 1.0 list from a mirror of the June release. If the 2.0 board does carry a Mathematics column, the clause is fine. If it does not, then selecting the latest release and setting the category to Mathematics names two things that cannot both be true, and the resolver has to choose.

The favourite survives either way, which is the reassuring half. Anthropic tops the Overall column on both versions and the Coding column on both. It is ranks two through five, and any Mathematics position, that the version answer moves.

Open any of the six boards and count the legs: thirty-two, of which eleven are not real. They are named Company A, B, C, D, E, F, G, H, I, K and Other. The alphabet skips J. Every one is quoted at 0.50 with zero volume, zero liquidity, a 0 bid and a 1 ask.

A 0 bid against a 1 ask has a midpoint of 0.50, and that midpoint is what gets published as a price. It is not a probability, it is an empty book describing itself. Eleven of them add a flat 5.50 to every naive sum.

So the raw sums are nonsense and the stripped sums are informative. October Overall: 643.1% raw, 93.1% stripped. October Mathematics: 650.5% and 100.45%. October Coding: 650.3% and 100.30%. November Overall: 661.5% and 111.50%. November Mathematics: 647.9% and 97.90%. November Coding: 666.5% raw and 116.50% stripped.

Two of those deserve a second look in opposite directions. October Overall strips to 93.1%, which is 6.9 points short of a complete book: there is unpriced probability there, most plausibly in the long tail of Chinese-lab legs quoted at 0.05% with no bid at all. November Coding strips to 116.50%, which is 16.5 points of overround, and the spreads in the next section show where it comes from.

The practical rule applies to every grouped Polymarket market, not just this series. Before comparing any leg to your own estimate, delete every leg with zero volume, zero liquidity and a 0-to-1 book, then re-add what is left. If the remainder does not land near 100%, the gap is telling you about the book rather than about the world.

A price of 89.40% on a board with a 5-for-5 record reads as settled. The hourly history says it is anything but.

In the 24 hours to 15:00 UTC on 24 October, the Anthropic leg on October Overall went 82.65% at 17:00, up to 95.70% by 21:00, held in the low nineties overnight, dropped to 76.05% at 06:00, and climbed back to 89.20% by 15:00. That is a 19.65 point range between 76.05% and 95.70%, inside one day, eight days from resolution, on $9,133.53 of volume in that leg.

The month is wilder. Over the 30 days to 24 October the same leg printed a low of 48.50% and a high of 98.70%. Google's October Overall leg is the better story: 1.20% on 24 September, a peak of 32.75% around 12 October, and 0.20% now. For about a week in mid-October this board had Google as a genuine one-in-three contender, and that view has since been marked down by a factor of over 160.

The November boards move for a different reason. Anthropic's November Overall leg opened at 25.00% on 1 October, reached 88.50% by 17 October and sits at 74.50%, down 14 points on the week. Tencent's leg opened at 13.15%, collapsed to 0.55% within two days, sat there three weeks and is now 10.55%, up 9.95 points on the week. Meituan did the same round trip from 32.20% to 0.55% and back to 5.90%. Those opening prints look like seeding noise rather than opinion, and the recent moves like the first genuine repricing the November boards have had.

None of this contradicts the streak. It does mean treating the headline as a stable consensus, on a board where a day's drift is wider than the whole gap between 93% and certainty, is a mistake the history is happy to punish.

Every price above is a midpoint. On the October boards that is close to fair; on the November boards it is frequently fiction, and the overround from the last section is its arithmetic shadow.

The October Overall favourite is 0.888 bid against 0.90 ask, a 1.2 point spread, and you can trade near the quoted 89.40%. The OpenAI leg on the same board is 0.001 bid against 0.014 ask: the ask is fourteen times the bid, so the published 0.75% is the midpoint of a book whose two sides disagree by a factor of fourteen.

November is worse everywhere. Moonshot's November Coding leg shows 10.75% off a 0.006 bid and a 0.209 ask, a 20.3 point spread. Anthropic's leg on that board shows 80.50% off 0.73 against 0.88, a 15 point spread on the favourite itself. November Overall has Tencent at 10.55% off 0.056 against 0.155 (9.9 points) and Meituan at 5.90% off 0.003 against 0.115 (11.2 points). Nvidia's November Coding leg prints 5.25% off a 0.003 bid and a 0.102 ask, and the proposition that Nvidia tops LiveBench Coding in November is doing far less work in that number than the empty book is.

Tencent is the clearest case. Its November Overall price travelled from 4.40% to 16.05% and back to 10.55% within 24 hours while Gamma recorded zero dollars of 24-hour volume on the leg. Nothing traded. A market maker shifted quotes across a 9.9 point spread and the midpoint followed. An 11.65 point intraday swing on no trades is not information.

The depth is odd too. On October Overall the Anthropic leg carries $32,075.91 of lifetime volume against just $1,553.83 of resting liquidity, nearly 21 to 1, while the OpenAI leg at 0.75% carries $8,460.47 against $4,829.48. The long shot holds 3.11 times the resting depth of the favourite. If you need to exit before resolution, the book is thinnest exactly where the consensus is strongest.

First, strip the eleven dead legs before any arithmetic. Company A through K and Other quote an empty book at 0.50, and they are the difference between 666.5% and 116.5% on November Coding.

Second, trade the ask, not the midpoint. Half the November legs carry spreads of 10 to 20 points, and on a leg quoted 0.006 against 0.209 the entire apparent edge can live inside the spread.

Third, check which release the board will show at check time. The markets name the latest available LiveBench release, 1.0 was frozen on 25 June 2026, 2.0 is a different test with different categories, and the two Mathematics boards hold $74,770.53 against a category the 2.0 list does not appear to carry.

Fourth, size to the resting book, not the volume. $32,075.91 of lifetime turnover on the October Overall favourite sits over $1,553.83 of liquidity. Volume tells you a market existed; liquidity tells you whether you can leave.

Fifth, remember one snapshot decides it. These are not month-average markets. A leading score posted on 30 November changes the November boards completely and one posted on 2 December changes nothing, and on those boards the answer is public for about twelve hours before the book closes.

Sixth, separate the streak from the price. Five wins from five is the strongest prior available on any AI ranking market here, and 93.00% is also a price paying 7 cents on the dollar for eight days of release risk. Both are true at once, and deciding which matters more is the whole trade.

Everything here is market data and context, not financial advice. Odsage is an independent information site and is not affiliated with Polymarket, LiveBench, Anthropic, OpenAI, Google or any lab named above. Prices, volumes and liquidity were read from the Polymarket Gamma API and CLOB price history at 16:10 UTC on 24 October 2026, and leaderboard figures come from third-party trackers because livebench.ai serves no readable board to a fetcher; all of them will have moved since.