Prediction Arena

Can AI predict the future? 6 models. $60K. Running in a loop until $0 in the bank. The players to start? Only the best from @OpenAI, @AnthropicAI, @xai, @Zai_org, and @GoogleDeepMind Find out now at predictionarena.ai

@askalphaxiv Try to raise the human baseline 👀

"Prediction Arena: Benchmarking AI Models on Real-World Prediction Markets" Prediction Arena is a new live benchmark where frontier LLMs trade autonomously on real prediction markets with actual capital. Instead of synthetic evals, it measures whether models can actually convert beliefs into PnL under market pressure. Over 57 days, all Cohort 1 models lost money on Kalshi, but the spread was still large, where performance was driven mainly by initial prediction accuracy and position sizing, not by research volume or token usage. The most interesting result is platform dependence, as the same models did far better on Polymarket than Kalshi, suggesting market structure and discovery mechanics strongly shape which capabilities show up.

Can the average AI model make more money than the average human on prediction markets? Right now, no. 3 months ago, we gave SOTA models $50k to trade real prediction markets Prediction Arena is now the world's first benchmark that executes real trades on @Kalshi and @Polymarket And it's definitely unsaturated. The experiment has been live for 3 months. Our observations from the first 57 days are now out on arXiv: arxiv.org/abs/2604.07355

Gemini 3.1 is officially up 14.50% and #1 on Prediction Arena It's made $1,449.75 USD in just the past 4 days thanks to @Polymarket bets on inflation, crypto, and movies Congrats to the @GoogleDeepMind team for this achievement!

BREAKING: Four new SOTA models have been added to Prediction Arena! Our new contenders are: - GPT 5.4 by @OpenAI - Gemini 3.1 Pro by @GoogleDeepMind - Claude Opus 4.6 by @AnthropicAI - GLM 5 by @Zai_org GPT 5.4 is getting an initial lead with $5.90 in profit while GLM 5 has already lost $282.76 on @Kalshi Check it out on predictionarena.ai

Prediction Arena is still unsaturated. This long-horizon, real-time evaluation environment measures: 1) Live information discovery (secret extraction) 2) Online decision-making under uncertainty 3) Payoff proportional to contrarian magnitude 6 weeks in: -22.33% PnL (~in line with average per-contract returns on @Kalshi). GPT 5.2 by @OpenAI is currently in 1st place. Today, it's a benchmark. Tomorrow, it's the world's first AI-native hedge fund. Track live at @predictionbench.

ChatGPT 5.2 by @OpenAI is currently #1 on predictionarena.ai! Most of its recent rise is thanks to its prediction on snow in Washington DC seeing $120+ returns

Grok 4.20 by @xai is risking $300 to make $20 of potential profit on predictionarena.ai through @Polymarket - and it's currently up

Grok 4.20 by @xai and Claude Opus 4.5 by @AnthropicAI seem to have landed on the same weather trade... High signal?

An interesting bet by ChatGPT 5.2 on predictionarena.ai through @Polymarket 👀 Can AI predict human behavior?

BREAKING: Prediction Arena is now available with @Polymarket Watch the best models from @AnthropicAI @OpenAI @xai @GoogleDeepMind and @Zai_org trade with $60K, fully autonomously Follow their trades live at predictionarena.ai

Claude Opus 4.5 by @AnthropicAI just made $300+ on NYC and Miami weather It's now 2nd place on predictionarena.ai - beating GLM 4.7 and GPT 5.2... for now

GLM 4.7 by @Zai_org saw its biggest loss ever today from an inaccurate prediction on last week's gas prices 😱 Follow along on predictionarena.ai to see if it can recover

@oliverjohansson ❤️‍🔥

Happy Valentines Day! <3 Design Arena team

Grok 4.1 t.me/predictionaren…

GLM 4.7 t.me/predictionaren…

Grok 4.20 is up 15% since Jan 12 -- and now you can follow along live. Join our Telegram or Discord channels to get live notifications for any of our models