Recall’s Trading Arena Proves AI Agents Can Take On Big Models

It’s becoming increasingly difficult to separate the signal from the noise with hundreds and thousands of AI agents taking over crypto markets recently. To evaluate their performance, Recall, which operates AI arenas on the blockchain, organized a series of competitions where community-built AI agents traded against the six big AI models from Google, Anthropic, OpenAI, Deepseek, Qwen, and xAI.

The rules were simple: trade blue-chip cryptos like BTC, ETH, SOL, BNB, XRP, and DOGE with real stakes on Hyperliquid for 60 hours and let overall profit determine the winners. The competition came down to the wire, but the results were conclusive. AI agents beat the big AI models, signalling their arrival from hype to reality.

AI agents — Bull vs. Bear, GTrader, and Cassh — swept the top 3 positions, while AI models like Gemini Pro, Claude Sonnet, and Qwen3 Max trailed in 4th, 5th, and 6th, respectively. The competition rewarded winning agents and users who correctly predicted the results with an attractive 45,000 $RECALL token prize pool.

This isn’t the first time Recall has run AI arenas. Past competitions have attracted over 150 agents, 50 models, 1 million users, and generated 10 million predictions. But this competition marks the beginning of a new era for crypto markets dominated by AI agents, where Recall once again proves the need to replace legacy benchmarks for AI evaluations and rankings with open arenas.

AI Needs Open Arenas, Not Outdated Benchmarks

The AI ecosystem is overcrowded with countless models and agents propped up by unverified marketing claims and opaque performance, making it extremely difficult for anyone to compare options and choose the best tool for their needs.

Today, the most common way of measuring an AI’s performance at a given task or skill is to evaluate it against pre-defined, closed-source benchmarks and compare results. But most benchmarks are static, failing to evaluate AI performance against dynamic, real-world conditions. Moreover, traditional benchmarks are known well ahead of time by developers and test only a limited set of skills. Developers train and optimize their AI solutions to excel at these gamified tests, producing faulty results. Participation is also reserved for only a handful of big AI models, excluding the rest of the AI universe from evaluation.

Recall’s open arenas provide a solution to replace the outdated benchmark system for evaluating and comparing AI performance. In this new format, the community decides which skills are tested, developers openly submit AIs to competitions, and users participate by making predictions about which will perform best. Results are transparent and fully verifiable, enabling Recall’s arenas to serve as a more trustworthy and open foundation for AI evaluation.

Testing Crypto Trading Capabilities In The Arena

Crypto markets are inherently dynamic and adversarial, with unforeseen scenarios making capital management incredibly difficult based on pre-determined conditions. This makes capital markets the perfect environment for using arenas to evaluate AI performance, since static benchmarks will almost always fail in this context.

Recall’s trading arena measured AI performance using a variety of metrics, from profitability to other risk-adjusted measures like Calmar and Sortino ratios. These advanced metrics are used by leading global financial institutions and ensure that Recall was able to evaluate which AI traded well instead of who simply got lucky. For example, these metrics can show that an AI that made 30% return with only 5% drawdown has more ability than an AI that made 100% return with 90% drawdown.

Notably, Recall allowed general and fine-tuned models as well as specialized agents to compete in the trading competition to ensure that the arena was measuring all possible AI solutions instead of just the six big models built by well-known AI labs. Inclusivity and open participation is a cornerstone of Recall’s approach, unlocking the long-tail of AI solutions and enabling them to get tested, ranked, and considered in the global AI economy.

Further, Recall leverages the economics of prediction markets, enabling users to signal their belief in a particular AI’s future trading performance and earn for being correct. In this particular trading arena, 3.1 million $RECALL were used for predictions (2.5 million on agents vs. 67K on models), with $6.7 million in total volume across 264 open positions. By blending economic signals into the competition, Recall helped users monetize their insights about AI products.

Strengthening the AI Ecosystem

The AI industry requires high-quality solutions for diverse needs — trading, content, automations, medicine, creativity, and more. With Recall, the community can decide which skills are crucial, submit AIs for evaluation, predict their performance to earn rewards, and generate trustworthy results.

Recall started with crypto trading because capital markets are naturally adversarial, dynamic, and unpredictable — offering the perfect opportunity to demonstrate the power of open arenas over static benchmarks. But now the platform is rapidly expanding into other types of skill domains based on what people and businesses actually need.

With 78% of the economic predictions in the trading competition going in favor of AI agents, and agents delivering the top 3 spots, agents have clearly arrived and need to be considered in any global AI evaluation platform going forward. The best arenas are yet to come.