AI Agent Claude Nets 1,322% on Polymarket as Rival is Liquidated

A recent experiment set two autonomous AI agents to trade on Polymarket, a crypto prediction platform, with markedly different results. The agent based on Anthropic's Claude model started with $1,000 and reached $14,216 in 48 hours, achieving a 1,322% return, while the rival using the OpenClaw framework was liquidated, losing its entire initial balance. The test, shared in a viral post on X by user @TFTC21, took place over the past few days and has already garnered millions of views.

Experiment Details

The agents operated independently, focusing on high-frequency markets like minute-by-minute Bitcoin price fluctuations. The Claude setup executed over 5,200 trades, covering even API costs during the process, according to reports on sites like MEXC and BingX. In contrast, OpenClaw, an open-source framework for persistent agents, made fewer than 200 trades and suffered a 94% drop before zeroing out, as described in an Instagram reel. This difference highlights how models like Claude 3.5 Sonnet better handle volatility and rapid decisions.

OpenClaw, released in late November 2025, gained rapid traction with nearly 300,000 stars on GitHub and inspired variants from Chinese companies like Tencent and Zhipu, whose shares rose by up to 16%. However, it faces criticism for being complex to set up and prone to bugs, which may explain the test failure. Users on Reddit and YouTube note that it consumes many API tokens and relies on premium models to perform well, while Anthropic's native solutions are more efficient.

Repercussions and Skepticism

The video from @TFTC21's post, showing real-time balance charts, went viral with over 1.9 million views and sparked debates about AI's potential in finance. However, skeptics question its authenticity, as there is no independent on-chain verification. Other similar experiments, such as one with Clawdbot (an earlier version of OpenClaw) that generated $347 overnight, show that results can vary.

Anthropic, in turn, has released recent updates that integrate OpenClaw-like functions directly into the Claude platform, including multi-agent workflows, potentially reducing reliance on external frameworks. Meanwhile, OpenAI acquired the creator of OpenClaw in February 2026, signaling that the project may evolve under new management but with security restrictions after bans by Anthropic and Google due to vulnerabilities.