LMArena Leaderboard September 2026: Where the Top Text Score Stands

LMArena, which now calls itself Arena, ranks AI models by people's votes: two anonymous models answer the same prompt and the voter picks the better answer. In the text leaderboard's update dated 25 September 2026, the highest overall score with style control turned off was 1518. One live prediction on versus asks whether any AI model will reach a given overall Arena score by 31 December 2026, with answers from 1520 to 1700. This guide covers the September scores, how the score works, how the prediction settles and what a position can cost you.

The Short Answer

• Top score in September 2026. claude-opus-5.5-high from Anthropic led the Text Arena overall ranking: 1518 with style control turned off, 1509 in the default view.
• Seven answers, and several can be right. The answers are 1520, 1530, 1540, 1550, 1600, 1650 and 1700. A model reaching 1535 would make both 1520 and 1530 right.
• The view matters. The prediction uses the score with style control turned off, not the default view.
• You can lose the whole amount. A winning contract settles at 100p and a losing one at 0p, and a position that settles at 0p is lost in full.

LMArena, Chatbot Arena and Arena

According to its own About page, Arena (formerly LMArena) was created by researchers from UC Berkeley, and the platform was previously known as Chatbot Arena. In 2025 the team formed a company, Arena Intelligence Inc. The old lmarena.ai address now forwards to arena.ai.

How the Arena Score Works

In Arena's battle mode, two anonymous models answer your prompt and you vote for the answer you prefer. The names are revealed after the vote, and only votes cast while the models are anonymous count towards the rankings.

Arena turns the votes into scores with the Bradley-Terry model, a statistical method for paired comparisons that it likens to the Elo ratings used in chess. Next to each score the leaderboard shows a margin, such as ±12, and the number of votes. Arena's leaderboard policy says a model needs at least 1,000 votes, and typically more, before its score is listed.

What Style Control Changes

Voters judge how an answer looks as well as what it says. Arena's research found that style has a strong effect on scores, so it built a version that adjusts for answer length and for the headers, bold text and lists in an answer.

On the leaderboard page this is set under Adjustments. The default view uses Style Control; to see the scores without it, choose None. In the 25 September update the top score was 1509 with style control and 1518 without it, for the same model, and the order below the top differs too. The prediction settles on the view without style control.

Where the Top Text Scores Stood in September 2026

The top five in the Text Arena overall ranking with style control turned off, in the update dated 25 September 2026, which the page says drew on 8,528,723 votes across 409 models:

1. claude-opus-5.5-high (Anthropic): 1518, ±12, from 2,307 votes.
2. claude-fable-5.1-max (Anthropic): 1511, ±7.
3. claude-opus-5-max (Anthropic): 1506, ±5.
4. claude-opus-4-6-high (Anthropic): 1504, ±3.
5. claude-opus-5-high (Anthropic): 1504, ±4.

Anthropic models held the top six places in that view. The highest model from another lab was Google's gemini-3.8-flash-high at 1494, marked preliminary. Arena marks a score as preliminary when its votes were collected before the model's public release, until enough new votes come in after it.

According to Arena's leaderboard changelog, claude-opus-5.5-high joined the Text Arena leaderboard on 25 September 2026. It had far fewer votes than the other four models in the top five, and the widest margin. Its 1518 is 2 points below the lowest answer, 1520. Arena's policy describes ratings stabilising as votes build up, so a new score can still move up or down.

How This Prediction Settles

Every market on versus names the source that settles it before you trade. Here the rules name the Text Arena leaderboard and its overall score with style control turned off. Each number is a separate answer with its own price between 0p and 100p, the market's estimate of how likely a model is to reach that score in time. The cut off is 05:59 Central European time on 1 January 2027. If any model reaches at least the number before then, a contract on that answer settles at 100p; if none does, at 0p. The rules also cover what happens if the leaderboard is unavailable.

Because the answers are thresholds, more than one can be right:

• A top score of 1525 before the cut off would make only 1520 right.
• A top score of 1545 would make 1520, 1530 and 1540 right.
• No model reaching 1520 would make none of the answers right.

So a higher number can never be more likely than a lower one. Its price will usually be lower, and the prices do not have to add up to 100p.

A Worked Example

Say the 1530 answer trades at 30p. You put 600p into it, which buys 20 contracts. The transaction fee is 2% of the stake, added on top, so it comes to 12p and 612p leaves your balance. The fee is shown before you confirm.

• If a model reaches 1530 before the cut off: the 20 contracts settle at 100p each, 2,000p in total.
• If none does: they settle at 0p, and the 612p is gone.
• If you change your mind first: while the market is open you can sell at the current price. Say weeks pass without a new top model and the price falls to 15p. Selling the 20 contracts gives you 300p back: a loss, but a smaller one than 0p.

In every case, the most this position could cost you was the 612p you saw before confirming. There is no leverage and no margin call, so a market moving against you cannot leave you owing money.

What Can Move the Prices

Prices here tend to follow AI news: a major release, a new model added to the Text Arena leaderboard, or fresh votes that move a score. Arena records new models and methodology changes in its leaderboard changelog. As the year end gets closer without a model reaching a number, that answer's price tends to fall, because there is less time left for it to be right.

None of that decides the result: only the score on the view the rules name does. A price is only the market's estimate at a given moment, not a promise.

Checks Before You Open a Position

1. Read the rules first. They name the leaderboard, the view and the cut off.
2. Check the right view. Choose None under Adjustments; the default view shows different numbers.
3. Look at the margin and the votes. A model with a few thousand votes can move more than one with tens of thousands.
4. Decide what you are prepared to lose in full. The stake and the 2% fee, shown before you confirm, are the most the position can cost you.

Safer Play is built into the app: deposit limits, loss limits, reality checks, time out and self exclusion, alongside GAMSTOP. The safer play page explains each tool.

Where to Find This Prediction

versus is a prediction market app across politics, finance, sport, AI, crypto and entertainment. The current predictions include AI questions such as this one: open any market to see its question, the price of each answer and the source that settles it. How it works follows one position from opening to settlement, and when markets close explains how cut offs work.

versus is licensed by the UK Gambling Commission under account 101143, which you can check on the Commission's public register. Positions settle in ordinary money, not cryptocurrency, and customer money is held separately from company money.

The Bottom Line

In late September 2026, the top overall text score on the LMArena leaderboard, now Arena, was 1518 with style control turned off, 2 points below the lowest answer. Whether any model reaches 1520 or more by the end of 2026 depends on releases and votes nobody can know in advance, and any position can settle at 0p, so only put in what you are comfortable losing. Prediction markets are entertainment with real money at stake, not a way to make money.

18+ only. Trade responsibly. BeGambleAware.org and GAMSTOP.