I launched a swarm of 10 AI models to compete in finding the best trading strategy. The results were surprising—every time, the cheaper, open-source models outperformed the most expensive ones. Claude Opus 4.6, which costs 10 times more to run, never beat the S&P 500 across three separate experiments. I ran the test multiple times, not believing the outcome at first. Then I ran it again. And again. The pattern held. This was an AI “agent swarm”—a setup where multiple large language models are deployed simultaneously, each given the same objective: to create the most profitable and risk-efficient trading strategy. I used a multi-agent interface to spawn the models, allowing them to operate independently while sharing a common goal. Each AI acted like a quantitative researcher, developing its own approach. The process began with each model generating a research plan. They would identify market trends, select indicators, define entry and exit rules, backtest strategies, and refine their logic. Some focused on momentum, others on mean reversion or volatility regimes. They all used historical market data, but their methods varied widely. What emerged was a fascinating divergence. The high-cost models—those praised for their reasoning and fluency—often produced overly complex strategies. They overfit data, relied on speculative patterns, and struggled with real-world constraints like transaction costs and slippage. In contrast, the cheaper models—often dismissed as “spyware” or unreliable—tended to generate simpler, more robust strategies. They focused on clear, repeatable signals. They avoided overcomplication. They prioritized consistency over cleverness. The results were consistent across three independent runs. The top-performing models were all open-source or low-cost proprietary systems. The most expensive model, Claude Opus 4.6, never came close to matching the S&P 500, let alone beating it. This experiment revealed a key insight: in trading, simplicity often beats sophistication. The best models aren’t necessarily the most advanced—they’re the ones that avoid overfitting, respect market reality, and stick to clear rules. It also highlights a growing trend: the most effective AI tools in practical domains may not be the most expensive ones. Sometimes, the best performance comes not from raw compute power, but from disciplined design, lean architecture, and a focus on real-world outcomes. The takeaway? In the world of AI-driven trading, cost efficiency and strategic clarity can outperform hype and price.