OSOdsageAI markets, decoded
Back to blog

AI Models

Which Company Has the Best AI Model? 2026 Odds at 86%

Which company has the best AI model? Polymarket prices Google at 86% for year-end 2026, up from 34.5% on 28 October. We read all 26 legs and the rules.

which company has the best ai model12 min read

Key takeaways

  • Polymarket's year-end market prices Google at 86.0% last trade, Anthropic at 10.0% and OpenAI at 2.0%, read from the Gamma API at 16:05 UTC on 21 November 2026. Seven weeks ago those numbers were almost exactly reversed.
  • The reason is on the leaderboard, not in the labs' press releases: Google holds ranks one, two and three on the Arena text board as of 21 November, with Gemini 4 Carbon at 1527, Argon at 1520 and Neon at 1513. Anthropic's best model sits fifth at 1506, 21 points back.
  • The same question is priced at three horizons and the prices disagree in a consistent way: 97.3% for 30 November, 91.0% for 15 December, 86.0% for 31 December. That implies an 11.6-point chance someone takes the top rank off Google in the 31 days between.
  • The three markets point at three versions of the same leaderboard. The year-end contract still names lmarena.ai, a domain that 301-redirects to arena.ai after a January 2026 rebrand; the two newer contracts name the explicit no-style-control URL and add an AutoEval exclusion the year-end contract does not have.
  • Polymarket resolves on the board's plain Rank column, Kalshi resolves on Rank (UB) and then on vote count and release date. Kalshi prices the same outcome at 84.5% against Polymarket's 86.0%, on roughly five times the volume.

On this page

Most of what you can read on the question of which company has the best AI model stopped being true in October. One widely linked explainer is dated 8 June 2026 and still has Anthropic as a 57.5% favourite. Another, last updated on 24 September, leads with Claude at 71c on Kalshi. Of the three ranking pages we checked only one had refreshed to the live number, and the live number is Google at 86.0% — the lead changed hands on one specific day in October that neither stale page mentions.

That gap matters because the contract is not a vote on which lab you think is better. It reads one column of one leaderboard at one timestamp, and three separate Polymarket contracts on the same question point at three different versions of that leaderboard.

So we pulled the year-end event from Polymarket's Gamma API at 16:05 UTC on 21 November 2026, rebuilt the full daily price series for the Google and Anthropic legs from the CLOB history going back to 5 June, read the resolution text of all three horizon markets side by side, and cross-checked the whole thing against Kalshi's equivalent contract and the current state of the board itself. The order below:

1. Where the odds stand across all 26 legs.

2. Why Google holding three ranks matters more than Google holding one.

3. The flip, dated to the day, and what caused it.

4. Three horizons, three prices, and the risk between them.

5. Three contracts pointing at three different leaderboards.

6. Why Polymarket and Kalshi can resolve this differently.

7. The leaderboard Anthropic is still winning.

8. The legs you cannot buy and the one named wrong.

9. Five checks before you take a side.

Where the best AI model odds stand right now

Twenty-six legs, fifteen of them named companies, read at 16:05 UTC on 21 November 2026. Prices are last trade with the best bid and ask beside them, then lifetime volume, 30-day volume and resting liquidity.

Google: 86.0% last, 86.0c bid against 87.0c ask, $723,683 lifetime volume, $406,346 in 30 days, $33,550 of liquidity.

Anthropic: 10.0% last, 9.0c against 10.0c, $637,223 lifetime, $260,712 in 30 days, $45,873 of liquidity.

OpenAI: 2.0% last, 1.8c against 2.0c, $344,416 lifetime, $119,369 in 30 days, $38,090 of liquidity.

Meta: 0.8% last, $347,611 lifetime. xAI: 0.3%, $284,284 lifetime. Then a floor of ten 0.1% legs nobody trades as a live outcome: Alibaba, Z.ai, DeepSeek, Moonshot, Mistral, Meituan, Microsoft, Amazon, Baidu and ByteDance.

Read cumulatively, Google and Anthropic together are 96.0% of the book and the other thirteen named companies share 4.0%. The fifteen named legs sum to 100.1% at last trade and 98.1% at the best bid, so unlike most Polymarket ladders this one has essentially no overround to argue about — the gap is inside the spread.

The whole event has turned over $4,620,669 since it opened on 4 June 2026, $1,501,605 of that in the last 30 days, against $1,586,556 of liquidity. It is the most heavily traded AI-model question on the venue, and its favourite has the thinnest book on the ladder. We come back to that.

Google owns ranks one, two and three

Here is the part the aggregator pages leave out, and it is the whole explanation for an 86% price. This contract resolves on the company that owns the single highest-ranked model. It does not ask whether Google's best model beats Anthropic's best model. It asks whether anyone at all can get above every Google model on the board.

As of 21 November 2026, the Arena text board reads: Gemini 4 Carbon (High) first on 1527, Gemini 4 Argon (High) second on 1520, Gemini 4 Neon (High) third on 1513. Then Meta's Muse Wonder (Max) fourth on 1511, Anthropic's Claude Opus 5.5 (High) fifth on 1506, Claude Fable 5 (High) sixth on 1504, and Claude Fable 5.2 (Max) and Fable 5.1 (Max) seventh and eighth on 1503 each.

Count the gaps rather than the ranks. Anthropic's leading model is 21 points behind the leader and 7 points behind the third Google entry. Meta's single top-five model is 16 points behind. A challenger does not need one good release; it needs a release that clears 1527 and holds there against three incumbents that each have their own upgrade path.

That is the structural reason the price keeps grinding up rather than jumping. Each additional Google model near the top adds a layer a rival has to climb over, and Neon is the newest of the three, and the board we read flags it as a new entry, which is consistent with the price step in the series below.

The 30 October flip, dated to the day

The CLOB daily series makes the turn unambiguous, and it is a single step rather than a drift. On 28 October the Google leg closed at 34.5% and Anthropic at 55.5%. On 30 October — the next day the series carries a print — Google was 56.0% and Anthropic 33.5%. A 21.5-point move on one side and 22.0 on the other, overnight.

What landed in that window was Gemini 4 Carbon. Google DeepMind announced it on 28 October, two weeks after Argon, and it debuted straight into first place on the text board with a score of 1528, ahead of Argon on 1520 and Claude Opus 5.5 on 1509. Arena's own results arrived on 30 October. The market repriced the same day the board changed, not the day the model was announced, which is exactly what a contract that reads a leaderboard should do.

Before that it was genuinely close. Anthropic led from the June opening through most of October, and the two legs traded inside two points for a fortnight — 42.5% against 41.5% on 9 October, 41.5% against 42.5% on 14 October. Nothing there looked like a 76-point gap waiting to happen.

Since 30 October it has been one way: 56.0% on 30 October, 65.0% on 9 November, 71.5% on 11 November, 78.5% on 18 November, 86.5% at the intraday read on 21 November. Anthropic went 33.5%, 29.0%, 21.5%, 17.5%, 9.5% over the same marks. The second-biggest single step, 58.5% to 71.5% between 10 and 11 November, lines up with the next Gemini entry reaching the boards.

Three horizons, three different prices

Polymarket runs this question at more than one date, and the prices across them are the most useful thing on the venue that almost nobody quotes. All three read at the same moment on 21 November 2026.

End of November, 9 days out: Google 97.3% last, 97.1c bid against 97.3c ask. Anthropic 1.1%. OpenAI 1.0%. Event volume $1,803,423.

15 December, 24 days out: Google 91.0% last, 90.0c against 91.0c. Anthropic 8.0%. OpenAI 1.3%. Event volume $52,104.

31 December, 40 days out: Google 86.0% last. Anthropic 10.0%. OpenAI 2.0%. Event volume $4,620,669.

Divide one by another and you get the market's own view of how fast the top of the board can change hands. The implied chance Google still holds rank one on 31 December given it holds it on 30 November is 86.0 ÷ 97.3, or 88.4%. From 15 December to 31 December it is 94.5%. So traders are pricing roughly 11.6 points of challenger risk into the 31 days of December, and about half of that into the second fortnight.

That shape reads in both directions. The market expects no surprise in the next nine days, and 97.3% is close to the practical ceiling for a contract that still carries resolution-source risk. It also says nobody thinks this is settled: an 86% favourite on a 40-day horizon is a long way from the 99%+ you see on a fixed outcome.

Three contracts, three different leaderboards

This is the finding that changed how we read the three prices above, and we have not seen it anywhere else. The three contracts do not point at the same resolution source.

The year-end contract, created on 3 June 2026, names its resolution source as https://lmarena.ai/ and describes the table as the "Chatbot Arena LLM Leaderboard", with results taken from the Rank section of lmarena.ai/leaderboard/text with style control off.

The end-of-November and 15 December contracts, created on 18 September and 11 November, name https://arena.ai/leaderboard/text/overall-no-style-control, specify the Rank column under Text Arena | Overall with adjustments set to none and the view filtered for models, and add a clause the year-end contract does not carry: models marked AutoEval at check time are not considered, whether or not they display a rank or score.

Those are not cosmetic differences. LMArena was rebranded to Arena in January 2026, five months before the year-end contract was written, and lmarena.ai now returns a 301 redirect to arena.ai — we checked, and lmarena.ai/leaderboard/text redirects to arena.ai/leaderboard/text, which is the default view rather than the explicit no-style-control path the newer contracts name. The year-end contract still asks you to find the right toggle on a renamed site.

The practical risk is narrow but real: a model on the board under AutoEval, or a board state where the default and no-style-control views disagree, resolves the November contract cleanly and leaves the year-end one arguing about which table its 301 landed on. Trade the spread between horizons and you own that difference too.

Polymarket reads Rank, Kalshi reads Rank (UB)

Kalshi runs the same question as KXLLM1, "Best AI on Dec 31, 2026?", and it has been open since 1 January 2026 — five months longer than Polymarket's version. The prices are close and the rules are not.

Kalshi's Gemini leg is 84.5% last, 84.1c bid against 84.6c ask, on $3,764,079 of volume and $1,479,067 of open interest. Claude is 9.6% on $3,987,777. ChatGPT is 3.6% on $3,606,783, then Kimi 1.0%, Muse 0.9%, Grok 0.7%, Qwen 0.3% and Ernie 0.1%. The eight legs sum to 100.7%.

Against Polymarket that is Google 86.0% versus Gemini 84.5%, a 1.5-point cross-venue gap on an outcome with no ambiguity about who is leading, and OpenAI 2.0% versus ChatGPT 3.6% — nearly double. The Kalshi book is the deeper one by some distance: $3.76 million on the Gemini leg alone against $723,683 on all of Polymarket's Google leg.

The rules explain some of the gap. Kalshi breaks ties under Rank (UB), the board's upper-bound rank column, then by highest Arena score, then by most votes, then by earlier release date. Polymarket breaks them under the plain Rank column, then by unrounded Arena score, then alphabetically by company name as listed in the market group. Rank (UB) treats models whose confidence intervals overlap as tied where plain Rank orders them strictly, so with four models inside 16 points of each other the two venues can read the same board and order it differently.

Anthropic is still winning a leaderboard, just not this one

The thing that makes 10% look harsh is that Anthropic was not beaten across the board in October. It was beaten on the one board this contract reads.

On 28 October, the day Carbon took first place on the text board, Anthropic's Claude Opus 5.5 (Max) was ranked first on Arena's Agent Arena with a net improvement of 21.1% over baseline, with Carbon second at 16.2%. Reporting since then has Claude Fable 5.2 (Max) leading the agent board on net improvement, and Anthropic also holds the top spot on at least one independent intelligence index. On cost per task the comparison runs the other way: Arena put Carbon's median at $1.80 per task, around 72% below Opus 5.5.

None of that enters this market. The contract names the text arena, the text arena is a blind pairwise human-preference vote, and preference voting rewards different things from agentic task completion. A lab can be the obvious professional choice for building software and still sit fifth on a board that asks which answer a person liked better.

So the 10% is not a verdict on Anthropic's models. It is the chance an Anthropic model outranks three Gemini entries on one table on one afternoon, and conflating the two is how a 10% price starts to look like a mispricing.

The legs you cannot buy and the one named wrong

Three structural details are worth naming before you treat the 4.0% tail as a tradeable long shot.

First, the tail is partly unquotable. Alongside its fifteen named companies the year-end market carries ten blank placeholder legs labelled Company A through I and K, plus an Other leg. Every one of those shows a 0.0c bid, a 100.0c ask, zero volume and zero liquidity. There is no market in them. If a company that is not on the named list takes rank one in December, the outcome resolves somewhere you could not have bought a position in at any price.

Second, the named list is out of date in two ways. The year-end market, written in June, lists xAI. The company was folded into SpaceX in a deal that closed in February 2026 and rebranded to SpaceXAI around 6 July, and the two newer contracts list SpaceXAI, not xAI. Those newer contracts also name six companies the year-end market never did — Nvidia, Xiaomi, Tencent, MiniMax, StepFun and Thinky — so a December surprise from any of them has a named, priced leg on the November and 15 December boards and no leg at all on the year-end one.

Third, the liquidity is in the wrong place again, which is a pattern on these ladders. The deepest books on the year-end market are Microsoft at $168,341, Amazon at $164,542 and DeepSeek at $160,714 — all priced at 0.1%. The 86.0% favourite that decides the event has $33,550. Market makers are quoting size on outcomes that should expire worthless and standing back from the one that matters.

Five checks before you take a side

First, open the board before the price. The contract is a reading of one table, so the only question that decides it is who is at rank one at the check time. If Google still holds the top three, 86% is a statement about three models at once.

Second, check which table your contract names. If you are in the year-end market you are resolving off a source URL that redirects; if you are in the November or December market you are resolving off an explicit no-style-control view with AutoEval entries excluded. Read your own contract's wording rather than the one you read last week.

Third, price the horizon, not the outcome. The three dates imply 88.4% for Google holding from 30 November to 31 December. If your own view of release timing differs from that, the trade is the spread between two dates rather than a bet on the favourite.

Fourth, compare venues before you size. Kalshi's book on this question is roughly five times deeper and 1.5 points cheaper on the favourite, with a different tiebreaker. A 1.5-point gap is inside the cost of crossing two spreads, so treat it as information about depth rather than as an arbitrage.

Fifth, do not buy the tail because it looks cheap. Eleven of the twenty-six legs have no bid, no ask and no volume, and several companies that could plausibly produce a December surprise are absent from the year-end list. A 0.1% leg with $160,000 of resting liquidity and no trades is a quote, not a market.

FAQ

Frequently asked questions

Which company has the best AI model right now?

On the Arena text leaderboard read on 21 November 2026, Google does: Gemini 4 Carbon (High) is first on 1527, ahead of Gemini 4 Argon (High) on 1520 and Gemini 4 Neon (High) on 1513. Anthropic's Claude Opus 5.5 (High) is fifth on 1506. Polymarket prices Google at 86.0% to still hold rank one on 31 December 2026.

How does the Polymarket best AI model market resolve?

It resolves to the company that owns the highest-ranked model when the Arena text leaderboard is checked at 12:00 PM ET on the market's date, with style control off. Ties go to the higher Arena score including unrounded values, and if models are still tied, to alphabetical order of the company names as listed in the market group.

Why do some sites still show Anthropic as the favourite?

Because they have not been updated since the lead changed. Anthropic led this market from June until late October, and pages still ranking for the query carry prices from that window: one dated 8 June 2026 reads 57.5%, another last updated 24 September reads 71c on Kalshi. The Google leg passed Anthropic for good on 30 October 2026.

What caused the odds to flip in October 2026?

Gemini 4 Carbon. Google DeepMind announced it on 28 October and it debuted at rank one on the text board with 1528, with Arena's results arriving on 30 October. The Polymarket Google leg went from 34.5% on 28 October to 56.0% on 30 October while Anthropic fell from 55.5% to 33.5%.

Is the Polymarket or Kalshi price better on this question?

They are close but not identical: Polymarket's Google leg is 86.0% and Kalshi's Gemini leg is 84.5%, and Kalshi's book is far deeper at $3.76 million on that leg alone. The rules differ too — Kalshi breaks ties on Rank (UB), then Arena score, then votes, then release date, where Polymarket uses the plain Rank column.

Could Anthropic still win the best AI model market for 2026?

The market says 10.0%, and the obstacle is arithmetic rather than opinion: an Anthropic model would need to clear 1527 and outrank three Gemini entries before 31 December. Anthropic does lead Arena's agent leaderboard on net improvement, but this contract reads the text board only.

What happens if a company not listed in the market takes rank one?

The year-end market carries ten blank Company A to K legs and an Other leg, all with no bid, no ask and zero volume, so that outcome resolves in a leg nobody could trade. Six companies the newer November and December contracts name — Nvidia, Xiaomi, Tencent, MiniMax, StepFun and Thinky — are not on the year-end list at all.

Sources and methodology

Sources used for this guide

Odsage combines public source links with prediction-market context, related market pages, calculator workflows, and visible FAQ content. Market prices are informational and are not financial advice.

Polymarket Gamma API - the snapshot every year-end price, spread, volume and liquidity figure here was read from at 16:05 UTC on 21 November 2026, plus the resolution text and 26-leg listhttps://gamma-api.polymarket.com/events?slug=which-company-has-best-ai-model-end-of-2026Polymarket - the live event page for the year-end contracthttps://polymarket.com/event/which-company-has-best-ai-model-end-of-2026Polymarket CLOB prices-history API - the Google leg's daily series, 166 points back to 5 June 2026, behind every dated price move herehttps://clob.polymarket.com/prices-history?market=46179164836162469124070224186387483861653666680647006036020478594822732769561&interval=max&fidelity=1440Kalshi public markets API, series KXLLM1 - the Best AI on Dec 31 2026 prices, volumes, open interest and Rank (UB) tiebreaker wording, read 21 November 2026. Kalshi's web page rate-limited our checks with HTTP 429, so the API is citedhttps://api.elections.kalshi.com/trade-api/v2/markets?series_ticker=KXLLM1&status=open&limit=100Benchleader, Arena Text leaderboard - rank, model, organisation and score for the top eight, stated as data at 21 November 2026, last modified 13:06 UTC that dayhttps://www.benchleader.com/benchmarks/lmarena_textOfficeChai, 28 October 2026 - the debut board (Carbon 1528, Argon 1520, Opus 5.5 1509) and the Agent Arena order, Opus 5.5 first at +21.1% against Carbon's +16.2%https://officechai.com/ai/gemini-4-carbon-tops-text-arena-places-2nd-on-agent-arena/Latent Space AI News, 28 October 2026 - Carbon announced two weeks after Argon, its 1528 text-arena score, the 30 October arrival of Arena results, and the cost comparison with Opus 5.5https://latent.space/p/ainews-gemini-4-carbon/Wikipedia, Arena (AI platform) - the January 2026 LMArena to Arena rebrand and the move to arena.ai, which is why the year-end contract's URL redirectshttps://en.wikipedia.org/wiki/Arena_(AI_platform)Lines, 8 June 2026 - the dated explainer still carrying Anthropic at a 57.5% implied probabilityhttps://www.lines.com/prediction-markets/tech/which-company-has-best-ai-model-end-of-2026Covers, last updated 24 September 2026 - Claude leading the Kalshi best-AI market at 71chttps://www.covers.com/entertainment/best-ai-prediction-markets