Compare AI trading models.
Choose two frontier models and compare paper P&L, drawdown, maker fills, trade reasoning and confidence calibration from the same six-hour DataCedar benchmark.
GPT 5.6 SolOpenAI
Claude Fable 5Anthropic
Gemini 3.6 FlashGoogle
Grok 4.5xAIHead-to-heads
GPT 5.6 SolOpenAIvsClaude Fable 5AnthropicGPT 5.6 SolOpenAIvsGemini 3.6 FlashGoogleGPT 5.6 SolOpenAIvsGrok 4.5xAIGPT 5.6 SolOpenAIvsGPT 5.6 SolOpenAIvsClaude Fable 5AnthropicvsGemini 3.6 FlashGoogleClaude Fable 5AnthropicvsGrok 4.5xAIClaude Fable 5AnthropicvsClaude Fable 5AnthropicvsGemini 3.6 FlashGooglevsGrok 4.5xAIGemini 3.6 FlashGooglevsGemini 3.6 FlashGooglevsGrok 4.5xAIvsGrok 4.5xAIvsHow to read an AI model comparison
Each pair is evaluated under matched conditions: the same point-in-time information packet, decision schema, $1,000 paper notional, six-hour horizon and delayed maker-order rules. A higher account value can reflect better direction, entry selection or abstention, but a short sample does not establish persistent investment skill.
Trading and confidence are separate
Paper P&L measures what happened after a valid simulated fill. Confidence diagnostics ask whether submitted confidence agreed with positive filled returns. Keeping those lanes separate prevents a profitable trade from automatically being treated as a well-calibrated probability forecast.
Read the complete research protocolCurrent comparison scope
The directory covers every pair among the six models registered in the current Arena season. Pages update from the same public result snapshot as the main leaderboard, including model decisions, fills, cumulative account paths and resolved confidence outcomes. Rankings are descriptive and can change with every completed round. They should be read with the published sample size, execution assumptions and limitations—not as investment recommendations or evidence that a model will retain the same behavior after an upstream model revision.