OpenAI indexes its financial data providers in-house to make every number clickable

The buildable half of this product is the half a large bank has already half-built. Templates carrying house style are an internal project. So is a rule about which desk may open which deal file. What a firm cannot stand up in a quarter is a licensed, indexed copy of five data providers' output sitting on the vendor's own hardware. Four of the five appear on the launch list; the fifth, Quartr, rests on reporting around the release ^[11][15].

OpenAI says hosting and indexing that data itself, instead of calling out to each provider on every request, improves retrieval speed, latency and how reliably it can trace a generated claim back to its source ^[12]. The retrieval reason is stable identifiers. An index you control gives every chunk an id that still resolves when a reviewer clicks it a month later, while five live vendor APIs put five availability windows and five rate limits in the answer path. The trade-off is freshness, and the write-up does not say how often the index refreshes.

That choice buys citations. The product surfaces granular citations so a user can check a number against where it came from while they work ^[13]. According to the dev.to account, the failure being designed against is a general chat model with no persistent access to properly entitled data and no awareness of a firm's house style: the output reads well and fails compliance review ^[14].

OpenAI publishes two benchmark numbers with the model, and both are claims about someone else's documents. OpenAI reports Astra about ten points above the previous model on OfficeQA Pro, which it describes as a rough proxy for office and financial document work ^[6]. For that to predict anything on an equity research desk, the desk's files and its questions would have to look like the benchmark's. On agentic, terminal-style tests, the same account says Astra beat its predecessor and Anthropic's comparable model at the time, at a lower cost per task ^[7]. That result describes OpenAI's bench that week, since the competitor is identified only as the comparable model of that moment.

The API exposes a reasoning effort setting with levels running from low up to max ^[5]. The analyst in the chat window has no equivalent dial; the setting sits with the team building on the API. That is where per-call spend gets decided, because summarizing a memo and building a comp table with sourcing should not cost the same.

The dev.to post paraphrases Nick Turley, OpenAI's VP of product, as describing the work as teaching ChatGPT to research like an analyst and back up its conclusions the way an analyst would ^[10].