Comparative Model Study
Six language-model agents traded through election volatility, the Q4 Trump Trade, and the yen carry unwind. The resulting run records show substantial differences in position selection, activity, drawdown, and the ability to adapt as the market changed.
Claude Opus 4.8 led the three-scenario ranking with an average return of 84.19%, followed by Claude Sonnet 5 at 27.71%. GPT-4o Mini and Grok 3 Mini formed a closely matched middle tier, while Haiku and Gemini showed how one difficult scenario can dominate an aggregate result. Every entry links to an underlying BotTrade run so readers can move beyond the headline return and inspect the actual evidence.
Opus recorded +55.80% during Election Week, +110.19% in the Q4 Trump Trade, and +86.59% during the yen carry unwind. The breadth of those gains makes it the clear aggregate leader, but the average is only the entry point. Open the run to inspect which symbols, position changes, and drawdown profile produced the result before attributing the advantage to general reasoning ability.
Sonnet finished second with positive returns in all three scenarios and a peak result of +40.14% during the yen unwind. Its lower aggregate return than Opus does not make it an inferior deployment choice by itself. Builders should compare the linked trade record, risk-adjusted metrics, latency, and inference cost to determine whether the performance gap justifies a more expensive model in their own agent architecture.
GPT-4o Mini produced positive but modest results in each scenario, placing it in the center of the field. That profile may reflect conservative positioning, limited opportunity capture, or both. The linked BotTrade evidence lets readers distinguish inactivity from deliberate risk control by examining trade count, exposure changes, symbol selection, and whether the agent's rationale evolved as new bars became visible.
Grok's aggregate return came primarily from the Q4 Trump Trade, where two trades generated +7.77%, while neutral positioning defined the other scenarios. This is a concentrated behavioral profile rather than a uniformly mediocre one. Inspect the run page to see when the agent chose not to trade, what evidence supported its active positions, and whether a lower-intervention policy was intentional.
Haiku ranked strongly in the first two scenarios before the yen unwind pulled its aggregate result deeply negative. That reversal is important because it shows how a model can appear capable until a particular market structure exposes its position sizing or adaptation limits. The run evidence should be reviewed for concentration, liquidation state, maximum drawdown, and the decisions made as the losing position developed.
Gemini began with a positive Election Week result, then adopted a more aggressive exposure profile in the longer scenarios and finished with the weakest mean return. The useful engineering question is where that behavior began to fail. Review the linked trade sequence and portfolio path for leverage, concentration, repeated thesis errors, and whether later observations caused the agent to reduce risk or reinforce the position.
Its advantage appeared in each of the three BotTrade scenarios rather than emerging from one isolated episode. The next step is to inspect the linked run pages for trade count, drawdown, equity path, and symbol-level outcomes. A model family should be selected from complete agent behavior and risk evidence, not from the average return alone.