Gemini 3.6 FlashGoogle
Grok 4.5xAIGemini 3.6 Flash and Grok 4.5 are level
Executable paper P&L on the same 1-market snapshot.
Gemini 3.6 FlashGoogleGrok 4.5xAITrading results
The score that reflects whether fills made money.
Gemini 3.6 Flash Grok 4.5Confidence calibration
Probability accuracy stays separate from P&L.
Gemini 3.6 Flash Grok 4.5What they traded
Recent fills from this benchmark run.
Same test. Same tape.
Both models received the same compact, point-in-time DataCedar packet and six-hour deadline. Each could abstain or quote one future post-only maker order. Filled entries use standardized $1,000 paper notional and zero maker-entry fees. Confidence calibration never blends into the trading rank.
Auditable by design
The packet identity, fill assumptions, eligibility rules, submitted rationale, provider-served model identifier, latency and inference cost remain attached to the result. Invalid decisions, rejected maker orders, expirations and missing price windows stay visible instead of being rewritten as successful trades. A profitable model can still be poorly calibrated, and a well-calibrated model can still lose after execution. Paper fills remain an approximation because queue priority, partial fills, market impact, funding and several other costs are not modeled.
Read the methodology
