Building an autonomous financial analyst with o1 and o3-mini

Endex is a small startup with a big claim: its AI system can do the kind of financial digging that usually eats up an analyst's entire afternoon, and it does it without a human nudging it through every step. The company built its product on top of OpenAI's o1 and o3-mini models, betting that reasoning-focused models, rather than plain chat models, are what finally make autonomous analysis workable.

The pitch is straightforward. Financial analysis is repetitive but demands judgment: pulling numbers from SEC filings, cross-checking them against earnings calls, spotting inconsistencies, and then stitching it all into something a portfolio manager can act on. Endex says o1's step-by-step reasoning lets its system plan out multi-part research tasks on its own, deciding what to check next based on what it just found, instead of following a rigid script written by an engineer.

o3-mini plays a different role. It's cheaper and faster than o1, so Endex uses it for the high-volume grunt work, things like scanning documents or running quick calculations, while saving the heavier reasoning model for the parts of the job that actually require judgment calls. That split matters for cost. Running o1 on every single query in an agentic pipeline gets expensive fast, and Endex needed something closer to production-viable economics if it wanted actual finance firms to pay for it rather than just marvel at a demo.

What's notable here isn't the idea of AI reading spreadsheets, plenty of tools already do that. It's the autonomy part. Endex's system is meant to chain together research steps with minimal human babysitting, which is the harder problem that most finance-AI products have quietly avoided. Whether it holds up on messy, real-world filings with contradictory numbers and management spin is the actual test, and that's the part OpenAI's blog post, understandably, glosses over.