October model odds put Google ahead of the arena check

The October model odds are a bet on a leaderboard, not on a product launch. On the Polymarket page for "Which company has the best AI model end of October?", the Google outcome row was labeled 64 percent when this page was opened on Oct. 2, 2026, with a Yes price of 64.1 cents. Anthropic's row was 37 percent. A second screen, CoinRithm, printed a vig-adjusted Google figure of 62.8 percent the same day. Those are not the same number, and this note keeps them apart.

TL;DR - Polymarket's loaded Google row read 64 percent, with Yes at 64.1 cents, while the header strip on the same page read 63.7 percent. Anthropic's row read 37 percent. CoinRithm's vig-adjusted Google price was 62.8 percent. - The market's total volume on the Polymarket page was $1,394,718. CoinRithm listed 24-hour volume of $333.4K and liquidity of $1.4M. Those are different clocks, not a correction of each other. - CoinRithm's reprint of the rules says the winner is the company behind the top model on arena.ai's Text Arena overall board at 12:00 p.m. ET on Oct. 31, 2026, with style control off and AutoEval models ignored.

What actually happened

This is not a new model. It is a market that was already open. Polymarket's page says the market opened Aug. 12, 2026, at 8:01 p.m. ET, and that it ends Oct. 31, 2026. CoinRithm labels the same window as Aug. 13, 2026, 12:02 a.m. UTC through Oct. 31, 2026, 4:00 p.m. UTC. An 8:01 p.m. Eastern time in August is midnight UTC the next calendar day, so the open stamps agree within about a minute. The CoinRithm end stamp matches a noon Eastern check on Oct. 31.

The page that loaded here was not one clean percent. A header strip put Google at 63.7 percent, Anthropic at 37 percent, OpenAI under 1 percent, and SpaceXAI under 1 percent. The outcome rows just under that strip were slightly different. Google showed 64 percent, Yes at 64.1 cents, No at 36.8 cents, with $295,798 of volume on that outcome. Anthropic showed 37 percent, Yes at 37 cents, No at 64 cents, and $346,546. OpenAI showed 1 percent, Yes at 0.6 cents, No at 99.5 cents, and $170,651.

The rest of the loaded book was a long tail. SpaceXAI was under 1 percent, Yes at 0.3 cents, with $138,857 on the outcome. Meta was under 1 percent, Yes at 0.2 cents, with $137,878. Alibaba, Moonshot, Baidu, Z.ai, Xiaomi, DeepSeek, Amazon, ByteDance, Nvidia, Thinky, MiniMax, Mistral, Meituan, Tencent, StepFun, and Microsoft were listed at under 1 percent, with Yes prices of 0.1 cents where a price was shown. The displayed percent and the cent price are not identical even on the leader. A 64.1 cent Yes is a 64.1 percent price. The row's large label said 64.

CoinRithm, opened the same day, said raw venue quotes summed above 100 percent and that it had normalized them. After that adjustment it priced Google at 62.8 percent, down 9.3 points over 24 hours, and Anthropic at 36.3 percent, up 15.5 points. OpenAI was 1 percent, up 0.3 points. It called Google the leader of 21 outcomes. It also printed a market-level 7-day change of minus 52.0 points, 24-hour volume of $333.4K, liquidity of $1.4M, and a 1.0 point spread. Do not subtract 62.8 from 64 and call the gap a price move. One figure is a raw row on Polymarket. The other is CoinRithm's vig-adjusted read of the same question.

The Polymarket page also carries an experimental AI-generated summary. The page says that summary is not trading advice and that it plays no role in how the market resolves. This article does not use that summary's percents. The October model odds in the header of this file come from the Google outcome row, the cell labeled 64 percent.

What the rules actually grade

CoinRithm reprints the venue terms and labels them informational. Under that reprint, the market resolves to the company that owns the model with the highest rank on the arena.ai Text Arena overall leaderboard when the table is checked on Oct. 31, 2026, at 12:00 p.m. ET. The specified view is the Rank column on the Text Arena overall page with style control off, adjustments set to none, filtered for models. The URL CoinRithm prints is the overall board with no style control.

Models marked AutoEval at that check are out, even if a rank or a score is showing. Models are ordered first by leaderboard rank. If two or more models are tied on rank, they are ordered by Arena score, including unrounded values under the leaderboard. If a tie still remains, alphabetical order of the company names in this market is the last breaker. CoinRithm's example is blunt: if two models were tied on the exact score, Google would rank ahead of SpaceXAI.

The October model odds do not grade a demo. They grade one filtered table. That is a narrow test. It is not "best" in a lab's own post. It is not a coding bench, an agent bench, or a speech bench. A model can be the one people quote all month and still lose if it is not first on that one table at that one hour. An AutoEval tag, if it is present at the check, removes a model no matter how high the number looks. A tie that survives the score comparison becomes a spelling contest between company names.

The Polymarket page showed a Rules heading, but the text that loaded here jumped from that heading to market context: volume of $1,394,718, an end date of Oct. 31, 2026, an open of Aug. 12, 2026, at 8:01 p.m. ET, and a resolver address beginning 0x69c47de9d. The rule sentences in the previous paragraphs are CoinRithm's reprint, not a second rulebook. If the two pages ever diverge, the resolver text on Polymarket is the one that pays. This article did not open arena.ai and did not record who leads that board on the day of writing.

What Polymarket thinks

The loaded leader is Google. On the outcome row that is 64 percent, beside a 64.1 cent Yes. The header strip's 63.7 percent is a third figure on the same site, and it is not the row. Anthropic is the clear second at 37 percent on its row, and that outcome shows more volume ($346,546) than Google's ($295,798). OpenAI's row is 1 percent. Everyone else on the loaded list is under 1 percent. Total volume on the Polymarket page was $1,394,718.

CoinRithm's adjusted book is the dissent inside the same question: Google 62.8 percent, Anthropic 36.3 percent, OpenAI 1 percent. The order of the top three matches. The tenths do not. A reader who treats every October model odds figure as interchangeable will mix three cells. A reader who screenshots only the header strip is not looking at the same cell as a reader who screenshots the 64 percent row, and neither of those cells is CoinRithm's 62.8.

For a different October company question, about agents rather than this arena table, see the October agent market. For a closed model that other notes use as a yardstick, see the Opus 5.5 note. Neither page resolves this market, and neither is a substitute for the noon check.

What to do as a reader (not a trade)

This is not investment advice, and it is not a reason to buy or sell a lab's shares. The October model odds move when traders move them. They are not a quality score you can cite as if you had run the eval.

If you repeat a percent, write down four cells first: the outcome name, the large percent, the Yes price in cents, and which site printed it. Then read the check time. Noon Eastern on Oct. 31 is not "sometime in October." Style control has to be off. AutoEval models do not count. A company can ship a loud model next week and still lose if that model is not the top eligible row at the check.

The two opened pages are Polymarket's best AI model market and CoinRithm's odds and rules reprint. Use the first for the live row. Use the second for the vig-adjusted figure and the arena instructions. Ignore any experimental blurb that the market page itself says does not resolve the contract.