Best AI Agent Odds 2026: Anthropic 91.5% on Polymarket

If you search for best AI agent odds you land on a Polymarket page showing Anthropic somewhere above 90% and very little explanation of what that number is attached to. The number is real. What it resolves against is not the thing most readers assume, and the cluster around it contains one contract whose favourite is priced at 82% while the company directly behind it cannot be traded at all.

There are five live agent contracts. We pulled all of them straight from the Polymarket Gamma API at 16:09 UTC on 12 September 2026, along with the arena.ai leaderboard they settle on, and every price below is a mid with the bid and ask printed beside it. Here is the cluster.

Best AI Agent, end of September: Anthropic 91.5% Yes, 91 bid against 92 ask. Event volume $216,667, book liquidity $71,977 across 18 priced legs.

Second-best AI Agent Lab, September: OpenAI 95.0%, 93 bid against 97 ask. Volume $57,577, book liquidity $12,343.

Third-best AI Agent Lab, September: Moonshot 82.3%, 80.8 bid against 83.8 ask. Volume $46,225, book liquidity $18,598.

Second-best AI Agent Lab, October: OpenAI 76.5%, 76 bid against 77 ask. Volume $34,511, book liquidity $14,989.

Third-best AI Agent Lab, October: Moonshot 49.5%, 49 bid against 50 ask. Volume $16,175, book liquidity $8,571.

All five check on 30 September or 31 October at 12:00 PM ET and expire at 03:59 UTC the following day. Put together they hold about $371,000 of lifetime volume, which is roughly a ninth of the single best-AI-model market for the same month. That size difference shows up in the spreads, and it is where the interesting pricing lives.

This is the part no competitor page mentions, and it is the difference between a correct position and an accidental one. The three contracts are not three views of one ranking. They read two different tables on the same leaderboard.

The Best AI Agent rules say resolution follows the company owning the model with the highest rank on the Agent Arena leaderboard when the table is checked filtered for Models. So it asks a question about a single model. Anthropic wins it if any one Anthropic model sits at rank 1, and a company with a deep bench of good-but-not-best models gets nothing.

The second- and third-best AI Agent Lab rules point at the same board filtered for Labs, where each company is ranked by its own best model. They also carry an exclusion the model market does not: models marked AutoEval at the check time are not considered, whether or not they display a rank or score.

Most of the time the two tables agree at the top, because a lab's rank is driven by its best model. They come apart in the tail. A lab can hold ranks 3 through 9 on the Models table and still be lab rank 3, while a lab with one strong model and nothing else can outrank it. If you are pricing the lower rungs off the Models view you are pricing the wrong table.

There is one more wrinkle worth reading before you size anything. The Lab markets say that if the lab ranking is ambiguous or unavailable, the companies get ordered by a fallback, which hands a genuinely close finish to a resolution judgement rather than a number. On a board where ranks 4 and 5 are separated by 0.67 percentage points, that is not a remote scenario.

Best AI agent odds at 91.5% and best AI model odds at 96.4% look like two readings of the same dominance. They are not. The two boards give Anthropic very different cushions, and the 4.9-point price gap is the market pricing that correctly.

The Agent Arena Models table, which is what the agent contract settles on, currently reads: rank 1 Claude Fable 5.1 Max, Anthropic, 13.85% net improvement. Rank 2 GPT 6 Astra Max, OpenAI, 12.39%. Rank 3 Claude Opus 5 High, Anthropic, 11.06%. Rank 4 Claude Opus 5 Max, Anthropic, 10.80%. Rank 5 Claude Fable 5 High, Anthropic, 9.09%. Rank 6 Claude Opus 4.8 High, Anthropic, 7.57%. Rank 7 GPT 5.6 Sol xHigh, OpenAI, 7.51%. Rank 8 Kimi K3 Max, Moonshot, 6.46%.

So the margin at the top is 1.46 percentage points, from one OpenAI model to one Anthropic model, on a board tracking 43 models across 1,587,202 sessions.

Now compare the Text Arena Overall board with style control off, which is what the best-AI-model contract settles on. Anthropic holds five of the top six places: claude-fable-5.1-max at 1510, claude-opus-5-high at 1505, claude-opus-5-max at 1504, claude-opus-4-6-high at 1503 and claude-opus-4-6 at 1498. The first non-Anthropic model is gemini-3.8-flash-high at 1494, sixteen points back, with an error bar of plus or minus 9.

One board gives Anthropic a 16-point gap over the nearest rival house. The other gives it 1.46 points over a model OpenAI shipped recently. A single OpenAI checkpoint landing in the next 18 days flips the agent crown and leaves the text crown untouched. That asymmetry, not sentiment, is what 91.5% against 96.4% is describing.

This is the sharpest thing in the cluster and it took a side-by-side read of the board and the leg list to find. The third-best AI Agent Lab market offers 18 tradable companies plus an Other leg. The Agent Arena Labs table does not match that list.

Labs ranking against the September third-place price, company by company: Anthropic rank 1 at 13.85%, priced 2.6%. OpenAI rank 2 at 12.39%, priced 2.25%. Moonshot rank 3 at 6.46%, priced 82.3%. Tencent rank 4 at 5.23%, no leg. Z.ai rank 5 at 4.56%, priced 6.0%. Meta rank 6 at 4.23%, priced 1.1%. DeepSeek rank 7 at 4.14%, priced 0.9%. Alibaba rank 8 at 3.97%, priced 3.1%. Google rank 9 at 3.84%, priced 3.2%. SpaceXAI rank 10 at 3.83%, priced 4.2%.

Two things fall out of that. First, Tencent is the company one slot behind the 82.3% favourite and there is no contract for it. The only leg that could absorb a Tencent finish is Other, which is quoted 0 bid against 100 ask with zero volume, zero book liquidity and no trade history at all. The market's most obvious tail risk has no price.

Second, the book ranks the western labs above where the board does. Meta is lab rank 6 at 1.1% while Google at rank 9 prices 3.2% and SpaceXAI at rank 10 prices 4.2%. That is not a standings read, it is a release-expectation premium: traders are paying for a Gemini or Grok checkpoint arriving before 30 September rather than for the ladder as it stands. It may well be right. It is a different bet from the one the contract's title suggests.

The liquidity behind these quotes is thin enough that the inversion can persist. Moonshot, the favourite, quotes 80.8 bid against 83.8 ask, a 3-point spread on an 82.3% mid. Z.ai, the highest-priced outsider, quotes 3.8 bid against 8.2 ask: a 4.4-point spread on a 6.0% mid, which is 73% of the mid in round-trip cost. Nobody is arbitraging a ranking inversion at that price.

Multi-outcome markets should sum to 100 cents. The gap between what they actually sum to and 100 is the cost of the ladder, and in this cluster it ranges from trivial to punitive. We summed every priced leg three ways: at the mid, at the ask, and at the bid.

Best AI Model, September, 21 priced legs: mids sum to 1.0085, asks to 1.0280, bids to 0.9890.

Best AI Agent, September, 18 legs: mids 0.9900, asks 1.0070, bids 0.9730.

Second-best Lab, September: mids 1.0200, asks 1.0880, bids 0.9520.

Third-best Lab, September: mids 1.0890, asks 1.2550, bids 0.9230.

Second-best Lab, October: mids 1.0095, asks 1.1490, bids 0.8700.

Third-best Lab, October: mids 1.1445, asks 1.4600, bids 0.8290.

The Best AI Agent line is the one to sit with. Its mids sum to 0.9900, which reads as a free dollar for 99 cents. At the offer the same basket costs 1.0070. The apparent 1% edge is entirely an artefact of quoting mids on legs where there is no bid at all, and it disappears the moment you have to pay someone. This is the single most common way readers talk themselves into an arbitrage that was never there.

At the other end, buying every outcome of the October third-best-lab ladder at the offer costs $1.46 to win $1.00, a 46% overround. Selling every outcome at the bid raises 82.9 cents against a certain $1.00 liability. Between those two numbers there is a 63-cent band in which the market maker is fine and you are not. Compare the flagship model market at 1.0280 and 0.9890 and you can see exactly what $3.18m of volume buys you.

Prices in this cluster moved hard in the seven days to 12 September, and they moved as mirror images of each other, which is the clearest evidence that the two ladders are being priced by people who understand the link between them.

Anthropic to be best AI agent went 82.00% to 91.50%, up 9.50 points. OpenAI on the same contract went 16.55% to 6.20%, down 10.35 points.

On the September second-best-lab market, OpenAI went 86.00% to 95.00%, up 9.00 points, while Anthropic went 13.50% to 3.60%, down 9.90 points.

On the much larger best-AI-model market, Anthropic went 86.00% to 96.40%, up 10.40 points, and OpenAI collapsed from 10.15% to 0.85%, down 9.30 points.

Read those together and the mechanism is obvious. Anthropic at the top of a board means OpenAI is second, so Anthropic gaining on the crown must show up as OpenAI gaining on second place by close to the same amount. It did, to within half a point on both boards. Coherent ladders look like this.

The proximate cause is Claude Fable 5.1 Max taking rank 1 on both the Text Arena and the Agent Arena boards. One checkpoint moved three markets in the same week, which is the thing to remember about every leaderboard contract on this site: they are not company-strength bets, they are bets on which single model is sitting at the top of a specific table on a specific afternoon.

The October contracts are not a longer-dated version of the September ones. They price a board that traders expect to have churned, and the size of that expected churn is worth putting a number on.

OpenAI to be second-best lab falls from 95.0% in September to 76.5% in October, an 18.5-point haircut for one extra month. Moonshot to be third-best lab falls from 82.3% to 49.5%, a 32.8-point haircut. Those are not small adjustments at the edges. The market is saying there is roughly a one-in-two chance that the third slot on the Agent Arena labs table belongs to somebody else by 31 October.

Who benefits is spread thin. On the October third-place ladder, Alibaba prices 10.5%, SpaceXAI 10.3%, OpenAI 9.5% and Google 6.6%, with the rest under 5%. Note OpenAI appearing at 9.5% for third place in October while priced 76.5% for second: the book is pricing a real chance OpenAI slips two rungs, which requires two labs to pass it.

The cost of expressing any of that is the problem. The October third-place ladder's asks sum to 1.4600 and its total book liquidity across all 18 legs is $8,571. SpaceXAI on that ladder quotes 3.7 bid against 16.9 ask, a 13.2-point spread on a 10.3% mid. You can have the view. Getting paid for it is another matter.

Moonshot's own drift is mild by comparison and points the other way: 44.00% to 49.50% over the past week, up 5.50 points, as Kimi K3 Max held rank 8 on the models board. The October market is repricing toward the incumbent, slowly.

Four things, in the order I would check them, and none of them involve a view on which lab is better at agents.

Check the filter before the price. Open the leaderboard, switch between Models and Labs, and confirm which table your contract names. If the position only makes sense on one of the two views, you have found the mistake before you funded it.

Price the tail you cannot trade. The September third-place market has no Tencent leg and a dead Other leg, so a Tencent finish resolves against every tradable outcome. Any position on that ladder is short an event with no quote, and the favourite's 82.3% should be read as 82.3% of a payout that also depends on the leg list being complete.

Sum the asks before you size. A ladder whose offers total 1.46 is charging 46 cents per dollar of certainty. Two of the six markets in this cluster clear 1.25. If you are taking the offer on the favourite in a ladder that expensive, the overround is coming out of your side.

Separate the standings bet from the release bet. Google at lab rank 9 pricing above Meta at rank 6 is a bet on a checkpoint shipping inside 18 days, not a bet on the current board. Those are different edges with different research and they should not share a position size.

The honest summary of best AI agent odds today: the cross-ladder pricing is tight and moved coherently all week, the favourites are roughly where the board says they should be, and the inefficiency that remains sits in the tail legs, the missing legs and the overround rather than in the headline numbers. That is a less exciting conclusion than an arbitrage, and it is the one the data supports.