Bloomberg

The stated mission is to reimagine the Terminal for AI, and without models the Terminal, data feeds and order management systems all remain. Above that base sit a purpose built financial language model, summarization products and an agentic query interface, a large layer on a platform that does not depend on it.

Shipped products summarize, answer and surface, and an analyst or portfolio manager acts on the result, with the agentic interface built for plain language queries rather than execution. The firm's own research analysts helped train the models against internal guardrail systems, though no runtime constraint, approval gate or statement of what the agent will not do is published.

A full technical paper documents the financial language model: 50.6 billion parameters, a 709 billion token corpus with its proprietary and public split, the compute used and benchmark results, including tasks where a general purpose model beats it. The paper is a checkable preprint rather than a peer reviewed one.

Generative features carry no named client and no quantified outcome. Subscriber counts, document library size and news volume describe the underlying business rather than what inference delivers.

The training corpus is described numerically, with the share drawn from the firm's own financial archive against public sources stated, and that archive is company collected data rather than customer material. Whether Terminal queries, uploads or connected research feed model improvement is never stated, nor is any tenancy position for ingested customer content.

The document research tool invites customers to connect their own proprietary research and analyze it beside the vendor's content. What happens to that material, and how long it is kept, appears in no processing terms or privacy position.

The vendor's pages name no certification, audit type or trust portal. For a platform carrying customer research alongside market data, that leaves assurance to procurement.

Group entities run regulated trading venues and administer benchmarks under supervision, but those authorizations do not extend to the Terminal, data and analytics business. If generative output ever fed an administered benchmark or venue function, that line would become arguable.

How a summary of an earnings call or filing selects what to include shapes what a trader sees first, and no governance position covers that choice. Fairness or bias evaluation for the generative features is likewise unpublished.

Summaries of earnings calls, filings and news reach traders who adjust positions on them, so an error can hit a market decision within minutes. No liability position, accuracy warranty or remediation route covers generated output.

The firm built its own model and documented it in depth, showing that building need not mean silence. No document maps which model powers which feature, and the paper dates from 2023 while the agentic interface arrived in 2026.

Order and trade management systems serve as the book of record for buy side and sell side desks, and communications archiving supports supervisory retention. Market data feeds into customer infrastructure and a widely referenced security identifier scheme add further system of record positions.

Hosting model, region and residency are unpublished on the vendor's pages. Third party coverage says the language model runs only on the firm's own infrastructure, a containment claim that would matter if the vendor made it.

A per seat price near $32,000 a year is reported consistently by independent sources and is close to common knowledge in this market, though the vendor publishes none itself. Whether generative features come with the seat or cost extra is the question that matters here, and nothing answers it.

More than 325,000 Terminal subscribers span the buy side, sell side and corporate treasury, along with government users and media, worldwide. The content estate reaches over 200 million company documents, and generative features sit inside that same platform rather than a narrower segment.