Anthropic's Claude disrupts prediction markets as bots turn profits on high-frequency trading

By surpassing most human traders and turning small initial capital into enormous returns, artificial intelligence-driven automated trading systems are altering the rules of prediction markets.

The most obvious indication of this change was when, on the decentralized prediction platform Polymarket, a trading bot based on Anthropic’s Claude AI model converted $1,000 into $14,216 in just 48 hours.

The experiment also compared Claude against a competing setup utilizing the OpenClaw framework, an open-source autonomous AI agent system.

The results demonstrated a significant performance gap: while the Claude-powered arrangement achieved a 1,322% return, multiplying its starting balance more than 13 times, the OpenClaw-driven setup was liquidated, losing its entire balance within the same 48-hour period.

Why is the change so noticeable? Risk was probably managed substantially differently by the two systems.

OpenClaw is a framework that uses techniques created by individual developers, whereas Claude is a huge language model created by Anthropic. It’s highly likely that their methods for controlling losses and position sizing differed.

While its competitor fell, the Claude-based bot, which is thought to have operated on Claude 3.5 Sonnet, maintained its position.

Bots finding edges humans simply miss

But these eye-catching numbers are not even the most extreme examples on Polymarket. One bot reportedly turned $313 into $414,000 in a single month. It traded only Bitcoin, Ethereum, and Solana in 15-minute up/down markets, placing bets of $4,000 to $5,000 at a time with a reported win rate of 98%.

The bot’s edge had nothing to do with predicting price direction. Instead, it spotted a delay between prices on Polymarket and confirmed momentum on major exchanges like Binance and Coinbase. When the actual probability of an outcome was already around 85%, but Polymarket still showed 50/50 odds, the bot stepped in and bought the mispriced side repeatedly.

Results from other systems have been as impressive. Using probability models trained on news and social media data, one bot generated $2.2 million over the course of two months. In order to keep up to date, it continued to retrain itself and concentrated on contracts where the market did not accurately reflect the real-world probability.

In the 5-minute Bitcoin market on Polymarket, a different trader set up three bots:

1. Near resolution, the first made straightforward directional bets.
2. Prices were compared to Chainlink Oracle data in the second.
3. Sitting on both sides of a trade to collect fees, the third functioned as a market maker.

Other tactics that nearly always yield a small profit include front-running thin-liquidity orders and buying both sides of a contract when the combined prices drop below $1.

The fairness debate and human traders

It has proven difficult for human traders to compete. Data comparing humans and bots using comparable techniques showed that while computers cleared approximately $206,000 with win rates exceeding 85%, humans employing similar strategies made around $100,000.

Even when their core strategy was correct, humans often lost any advantage due to poor stake sizing, late admissions, and insufficient risk controls.

The prevalence of automated technologies has sparked discussions about fairness. Anthropic has made a strong statement opposing the use of its technology for autonomous weaponry and surveillance, and it has publicly positioned itself as an AI safety company.

Critics now argue that allowing bots to systematically outpace and drain human traders on prediction markets raises questions that are not so different. The tension of whether a company that cautions against eliminating human oversight in military contexts should allow it in financial markets remains unresolved.

In addition to fairness, the swift ascendancy of AI threatens to undermine a fundamental democratic principle of prediction markets: the aggregation of varied human assessments. This could lead to echo chambers filled with biases optimized by machines and exacerbate the divide between those who possess AI and those who do not.

Get seen where it counts. Advertise in Cryptopolitan Research and reach crypto’s sharpest investors and builders.