Claude Opus 5.5 tops paper-trading test at -0.0078%

The experiment is structured as a controlled comparison: every route sees the same market-price snapshot and the same fill rules, while each model receives only its own portfolio and recent decisions. That design makes account values comparable within the test, but a caveat applies: chat models return an action plus a short reason, while Jev 1.14 uses a structured Decisions API that displays a distribution across choices, so differences in the decision interface remain a limit on comparison. Execution is simulated against Coinbase Exchange public data, buying at the best ask and selling at the best bid, with no market impact, partial fills, queue position, or extra latency slippage reproduced.

The snapshot itself is modest. With 5 completed rounds and 23 simulated trades, the spread between first and last was narrow: Grok 4.6 sat last at $9,996.17, a -0.0383% return. The order shown follows the API response, and models with identical account values are effectively tied. Leading at one moment only means the account value was higher at that point in the simulation, not that the model has general capability, and earlier and later captures showed a different model in first place.

A real answer would require observing multiple market phases and comparing drawdown, exposure, fees, failures, and consistency alongside returns. It notes that similar efforts exist, such as Nof1's market research, the TradingAgents multi-agent framework, and the FinRobot financial agent platform, but their differing goals and execution methods mean their results cannot be placed directly into this ranking.