DeepSeek and Qwen Lead Live AI Crypto Trading Test as GPT 5 and Gemini Suffer Heavy Losses

A $10,000 contest produces sharply different results

DeepSeek Chat V3.1 had turned a $10,000 trading account into roughly $22,500 by Monday afternoon, October 27, 2025, taking the lead in a live cryptocurrency experiment that has left some of its biggest rivals nursing heavy losses. Its gain of about 125 percent put it ahead of Alibaba's Qwen 3 Max, which was up nearly 95 percent after briefly holding first place between October 24 and 26.

OpenAI's GPT 5 and Google DeepMind's Gemini 2.5 Pro were at the other end of the table, with losses of roughly 60 percent. xAI's Grok 4 and Anthropic's Claude 4.5 Sonnet had more modest gains, ranging from the low to middle tens in percentage terms.

The six models are competing in Alpha Arena, a challenge launched by the US research group Nof1. Each received the same starting balance to trade real money autonomously on Hyperliquid, a decentralized cryptocurrency exchange. Trading began on October 18 and continues until November 3.

The early results put two models based in China ahead of their Western competitors, but the route to the top has been turbulent. DeepSeek suffered a sharp fall in account value before recovering, while Qwen's large bets helped it take the lead and later contributed to its retreat into second place.

Alpha Arena makes those decisions unusually visible. Trades are recorded on chain, meaning they appear in the blockchain's transaction record, and a public leaderboard displays holdings, returns and each model's written explanation for entering or leaving a position. The experiment tests whether language models can make money while controlling risk in a market that changes quickly.

What the models are trading

Nof1 gives every participant the same system prompt, the instructions that define its task and trading rules, and the same market inputs. Each model then decides whether to buy, sell or hold contracts linked to bitcoin, ether, solana, dogecoin, BNB and XRP.

The contracts are perpetual swaps, a type of futures contract without an expiry date. Instead of ending on a fixed date, they use a funding rate to help keep their prices close to the underlying market. That rate periodically pays one side of the trade and charges the other. For a trading agent, it is another cost or potential source of income to weigh alongside price movements.

The models can use leverage, which lets them take positions larger than the money committed to a trade. A favorable price movement can produce a bigger gain, while an unfavorable movement can deepen a loss. The bots must set take profit and stop loss levels, price points intended to close a position after a desired gain or an unacceptable loss.

Alpha Arena evaluates returns adjusted for risk rather than profit alone. That approach considers how much uncertainty and potential loss a model accepts to earn its return. A large account balance is part of the picture, but the fluctuations along the way matter too.

Each model trades through a dedicated Hyperliquid wallet address that the public can monitor. The agents have no access to live news or proprietary data streams. This common restriction reduces their exposure to headline driven decisions, while also limiting their ability to react to sudden events outside the market information they receive.

How the leaderboard changed

DeepSeek's advance accelerated after the opening weekend. It moved from an early profit in single digits to growth exceeding 100 percent by the beginning of the second week. Qwen kept close through large bets on market direction and briefly moved ahead.

The main dates in the contest so far are:

• October 18: Six models began trading with $10,000 each.
• October 24 to 26: Qwen briefly overtook DeepSeek before returning to second place.
• October 27: By Monday afternoon, DeepSeek was up about 125 percent and Qwen nearly 95 percent, while GPT 5 and Gemini were down roughly 60 percent.
• November 3: The current season is scheduled to end.

GPT 5 and Gemini fell behind as their losses at times approached or exceeded half of their starting capital. Grok and Claude spent stretches near their starting balances or with small positive returns before showing the more modest gains recorded in Monday's update.

DeepSeek's leading position has also included severe swings. A community update tracking its wallet described an intraday fall in account equity of more than 40 percent during a sharp market downturn, followed by a recovery. Account equity is the value of the account including gains and losses on positions that remain open.

A rebound in bitcoin and a recovery in ether after weeks of choppy trading helped models willing to hold long positions, bets that prices would rise. Several agents shifted together from short positions, which benefit from falling prices, into long positions as sentiment improved. Their shared direction helps explain why returns sometimes moved together and why a sudden reversal could hurt several competitors at once.

DeepSeek spreads its bets and manages exits

DeepSeek has generally spread its positions across the six available assets, often using moderate leverage and defined exit levels. Its early trades included profits on ether, solana and bitcoin, alongside a small loss on XRP.

The model's public reasoning feed shows repeated entries, prompt exits from losing positions when its rules are breached, and patience with positions that are making money. Observers following the contest have credited its performance to discipline, balanced allocation of risk and active management rather than a succession of unusually large bets.

DeepSeek's diversification means its money is distributed across several assets instead of resting entirely on one coin. The assets can still move together, as the contest's synchronized shifts show. Its account has experienced sharp declines, but its stronger upward runs have so far outweighed those setbacks.

The lead therefore rests on both successful market direction and the way positions were handled. The public record allows followers to examine the model's sizing, entries and exits, alongside the return displayed at the top of the leaderboard.

Qwen takes a more concentrated route

Qwen's trading has featured large positions concentrated in individual assets. At one point it committed fully to bitcoin, a decision that worked during a rally and helped propel it toward the top of the contest.

Later, Qwen shifted to a large ether position with greater leverage, betting that the price would rise. A pullback hurt that trade and contributed to its move from first place back into second.

Qwen's approach makes its results especially sensitive to the direction and timing of a particular coin. When that direction is favorable, a concentrated position can deliver substantial gains. When prices move against it, the same concentration leaves more of the account exposed, and leverage increases the damage.

By October 27, its gain near 95 percent still placed it well ahead of the four other competitors. Its temporary lead between October 24 and 26 showed how quickly the order could change, even as DeepSeek and Qwen remained the strongest performers in the latest update.

Why the other four models struggled or lagged

Grok has shown flexible positioning and periods of steady wins, although it has also lost ground after promising stretches. Claude stood out by opening one of the largest single positions seen in the contest, taking substantial risk after waiting for a setup in which it appeared to have greater confidence.

GPT 5 began cautiously and then struggled to adjust as conditions changed. Gemini traded rapidly and frequently reversed its direction. Those changes contributed to larger losses when market volatility increased.

Outside analysis of the contest has pointed to a tendency among some Western models to surrender early gains and chase prices after they have moved. That behavior has been described as consistent with systems relying too heavily on patterns learned from past data rather than adapting effectively to live feedback.

Overfitting is the term used for this kind of weakness: a system becomes too closely suited to historical patterns and struggles when new conditions behave differently. Reviews of Alpha Arena's first week have raised concerns about adaptability and risk controls, particularly the use of leverage, the frequency of trading and the placement of stops.

The six models received identical inputs, yet their choices differed in how much money they put at risk, how many assets they held and when they acted. Execution, cash management and discipline around exits have helped separate the leaders from the trailing models.

What may explain the Chinese models' lead

DeepSeek's origins have attracted attention because it was spun out of High Flyer Quant, a hedge fund known for quantitative strategies. Quantitative trading uses data and mathematical methods to guide investment decisions.

Investors and researchers have speculated that DeepSeek's training included financial signals of high quality and reinforcement cycles directed toward market decisions. Reinforcement training involves feedback that encourages some actions over others. Those proposed explanations for its performance remain speculation about how the model was trained.

Training cost is another point of discussion. Technical materials cited by industry analysts put DeepSeek's full training cost in the range of a few million dollars, compared with the billions spent shaping general chatbots. The argument is that a focused training approach may help a model interpret market structure and funding dynamics effectively, even with a much smaller budget. Lower spending alone does not guarantee better trading.

Experienced quantitative traders following the experiment have also argued that detailed financial instructions can matter as much as model size. A prompt that prioritizes position sizing, cash reserves and strict stops can guide trading behavior. A model highly capable at coding or answering difficult questions may still struggle with the timing of entries and exits in a live order book, the changing collection of buy and sell orders in a market.

Alpha Arena's common prompt makes the models' interpretations of those instructions part of the test. DeepSeek's diversified positions and Qwen's concentrated bets illustrate how different strategies can emerge from the same starting rules and information.

A public test of decisions under pressure

Jay Azhang, Nof1's founder and the creator of Alpha Arena, sees financial markets as a setting for testing the limits of autonomous systems. He described that purpose in direct terms:

Financial markets are the best training environment for the next era of AI.

Nof1 says the aim is to examine how these systems operate under pressure in a changing, adversarial environment, rather than to name a permanent champion. Competitors must respond to other market participants while dealing with uncertain prices, costs and the consequences of their previous decisions.

The Alpha Arena team has acknowledged community concerns about front running publicly identifiable models. Front running involves acting ahead of an expected trade to benefit from the price movement it may cause. The organizers have not detailed a specific response beyond transparency and continuing scrutiny.

The reasoning published with trades gives outsiders a way to examine each model's stated intentions. Language models can still produce confident but incorrect explanations, often called hallucinations. In trading, such errors can become expensive when they feed into decisions without strict rules to contain them.

Correlation is another concern. Agents built on similar base models and given similar prompts can make similar choices. Those shared positions can become crowded trades, with many systems trying to buy or sell in the same direction. When the market reverses, losses can strike several agents together.

What investors can learn from the experiment

Some equity analysts have suggested that sustained outperformance with manageable declines could lead professional investors to reconsider the need for increasingly complex quantitative trading systems. Their argument depends on agents delivering durable results, not simply enjoying a profitable stretch.

Others point to the contest's restrictions. The bots cannot see live news and lack the proprietary data and infrastructure used by modern trading firms. Their performance is being measured within that particular setting, with six crypto assets and a common stream of market information.

A more limited use is also attracting interest: model explanations may give traders new ideas and a quicker way to test them, even if fully autonomous agents struggle to outperform experienced human teams across a market cycle.

The heavy losses in several accounts show how late reactions, poor position sizing and weak execution can damage a trading system. The leaders show that agents can also identify a favorable trend and stay with it. Both outcomes are appearing in public, with the decisions and account movements available for examination.

Venue operators and regulators have reasons to watch fairness, disclosure and compliance if autonomous agents grow into a larger presence in public markets. Alpha Arena's transparent format brings those concerns alongside the more immediate contest over returns.

Public wallets, prediction markets and the final week

Each bot's exclusive Hyperliquid wallet allows followers to monitor its activity. The dashboard brings together positions, realized gains from completed trades, unrealized gains or losses on open positions, and the latest model reasoning.

Prediction markets appeared around the outcome later in the first week. At one snapshot, Polymarket bettors gave DeepSeek the highest chance of finishing first, with odds of roughly four in ten and tens of thousands of dollars in trading volume.

The contest has also drawn comments from prominent cryptocurrency figures. Experienced exchange operators and fund managers have returned to the idea that a distinctive strategy can matter more than raw computing power when algorithmic competitors face each other.

Nof1 gathers the leaderboard, rules, model pages, wallet information, positions and commentary on the organizer's website. With trading continuing until November 3, the positions visible there can still change the final ranking.

The organizers plan to extend the format beyond cryptocurrencies into equities and other assets. A consumer platform for investing through agents is also on their roadmap, intended to let individuals test or deploy models with guardrails. No launch date was given for those plans.

Other participants in the wider community are experimenting with related challenges and open agents that anyone can ask to place a mock trade. Together with Alpha Arena's public records, these projects are creating opportunities to study autonomous market decisions as they happen.

The current season's lasting output may be its record of decisions under stress. Engineers and traders can use the same public benchmark to examine successful trades, costly mistakes and the handling of exits. As of October 27, DeepSeek and Qwen lead, but several days of live trading remain before the November 3 finish.

The Essentials

• Alpha Arena began on October 18, 2025, and ends on November 3, with six models trading real money on Hyperliquid.
• Each model started with $10,000, identical instructions and the same market inputs, without live news or proprietary data streams.
• By October 27, DeepSeek Chat V3.1 had about $22,500, a gain of roughly 125 percent, while Qwen 3 Max was up nearly 95 percent.
• GPT 5 and Gemini 2.5 Pro had losses near 60 percent. Grok 4 and Claude 4.5 Sonnet recorded more modest gains.
• DeepSeek generally spread its positions across assets, while Qwen made large concentrated bets on bitcoin and ether.
• Public wallets and dashboards show positions, returns and the models' written trading explanations.
• Nof1 plans to expand into other assets and develop a consumer platform for investing through agents.