An AI Trained on Wall Street’s Own Data Just Beat GPT

Eighty-four point seven percent. That's the accuracy rate a custom AI model built by Bridgewater Associates and Mira Murati's startup Thinking Machines Lab achieved on financial document tasks, beating out GPT, Claude and Gemini, and doing it at nearly 14 times lower cost.

The secret wasn't smarter AI. It was smarter data.

General-purpose models, trained on vast amounts of public financial information, still averaged just 50% accuracy when asked to do the routine work of investment analysis, things like flagging whether a central bank document signals a rate change, or where boilerplate ends in a 300-page filing.

The difference came from fine-tuning a model on Bridgewater's own expert-labeled data, encoding the kind of judgment that experienced analysts build over years but can rarely put into words.

Bridgewater is not alone. Mastercard has taken a similar approach with its transaction data. The lesson emerging across the financial sector is that the bottleneck isn't model capability. It's proprietary data, and who controls it.