Meta tests 'Arena', a Llama-powered prediction-market app, with employees

Meta is testing a new consumer app with an AI-generated core. Arena, now in internal testing with employees, is a prediction market in which users forecast outcomes with virtual tokens instead of real-money bets. That design choice keeps Meta clear of the gambling and commodities regulation that real-money platforms like Kalshi and Polymarket face.

Llama does the content work. Meta's foundation models continuously generate forecasting questions, a job that would otherwise need human editors to keep up with the news, and they personalize the ranking of each user's feed. In effect, the app turns a language model into a constant source of 'what will happen next?' prompts tuned to each user's interests. The stated ambition is large: up to 100 million monthly active users in the 18-to-34 demographic, the group Meta has struggled to hold on Facebook.

The experiment comes amid a wider debate about Meta's AI direction. This week Meta also released Muse Ember 1.0, a low-cost hosted model with no published weights, and Muse models arrived on Google Cloud. Community observers read this as Meta stepping back from the open-weight Llama strategy, so it is notable that Arena keeps Llama inside a consumer product. Prediction markets have also become a hot category, and Meta, as usual, is trying to bring a trending format into its social graph.

Caveats: this is an internal test reported by secondary outlets, not a launch, and Meta kills many such experiments. Questions to watch include how Meta will moderate AI-generated questions on sensitive topics such as elections, violence or individuals' health; whether virtual tokens will ever be convertible to anything of value; and how regulators see a gamified forecasting app aimed at young adults. The 100-million-MAU goal is ambitious for a stand-alone app.