Bloomberg

The Clearwater band, consistent with the other data majors, and the vendor's own framing supports it: the stated mission is to reimagine an existing Terminal for the age of AI rather than to sell inference as the product. Strip the models and the Terminal, the data feeds, the news operation, the order management systems and the messaging network all remain.

What sits above that base is unusually substantial for a C, spanning a purpose built financial language model, three shipped summarisation products, a document research tool and an agentic query interface.

The shipped products summarise, answer and surface, and an analyst or portfolio manager acts on the result, with the agentic interface positioned as a way to query data in plain language rather than to execute. The company also states that its own research analysts were used to train the generative models against internal guardrail systems, which is a real description of how the models were shaped rather than an assertion that they are safe.

Held off A because those guardrails are described as applied during training rather than as published runtime constraints, and no approval gate, confidence threshold or statement of what the agent will not do was located.

The most thoroughly documented model in this index by a wide margin, and the only A on this axis earned from a formal technical publication. The company released a full paper describing its financial language model: 50.6 billion parameters, 70 transformer layers, 40 attention heads, a dimension of 7,680, a 709 billion token corpus with the proprietary and public split stated to the percentage point, and the compute used down to accelerator count and training duration.

Crucially it also published benchmark evaluations against named public financial reasoning and entity recognition tasks, including results on which a general purpose model outperforms its own. Publishing an evaluation that makes your model look worse is the strongest form of this disclosure and essentially nothing else in the index does it.

One distinction preserved for accuracy: the paper sits on a public preprint archive rather than having been peer reviewed, so it is checkable and falsifiable rather than externally validated. Held at A regardless, with the caveat that the paper is from 2023 and no document maps which model now serves which shipped feature.

The vendor names no client and publishes no quantified outcome attributable to the generative features. The large figures available describe the platform rather than the inference: subscriber counts, document library size and news volume are scale claims about the underlying business. Third party estimates of generative feature adoption exist but come from low quality aggregator content and are not credited under the source test.

The training corpus is described with a specificity no other vendor here approaches: the proportion drawn from the firm's own accumulated financial archive against public sources is stated numerically, and the archive itself is the company's own collected market, news and document data rather than customer material. That answers the question most vendors leave open.

Held off A because nothing states whether Terminal user queries, uploaded documents or connected customer research feed model training or improvement, no retention terms are published, and no tenancy or isolation position is offered for the customer content the document tool is designed to ingest.

The vendor publishes no processing terms, retention schedule or privacy position for the generative features in the material reviewed. The gap becomes pointed with the document research tool, which is designed to let a customer connect its own internal proprietary research and analyse it alongside the vendor's content, and nothing published addresses what happens to that material.

The vendor names no certification, audit type or trust portal in the material reviewed. Recorded as an unverified absence rather than an evidenced one, since no trust or security page was reached in this pass, and queued alongside the other data majors built in the same session.

Graded C on the division scoping principle, the sixth application across recent sessions. Group entities operate regulated trading venues and administer benchmarks under supervision, which are genuine authorisations, but this record is scoped to the Terminal, data and analytics business and neither reads across to unregulated software and content. Flagged as a live check on the same basis as the peers: if a generative output feeds an administered benchmark or a regulated venue function, read across becomes arguable.

The vendor publishes no fairness or bias evaluation for the generative features in the material reviewed, and no governance position for how a summary selects what to include. Flagged as the most likely of this vendor's C grades to be wrong: the published technical paper is the obvious place for toxicity, bias and safety evaluation to appear, papers of that kind routinely carry such sections, and it was not read in full during this pass. Queued as a named check.

The vendor publishes no liability position, indemnity, accuracy warranty or remediation route for generated output. The exposure is direct: summaries of earnings calls, filings and news are consumed by traders and portfolio managers who adjust positions on them, and a summarisation error reaches a market decision within minutes rather than through a review cycle.

The clean counterexample to a pattern this index has recorded repeatedly, and it deserves to be flagged as such. The standing observation has been that vendors who license capability name their suppliers while vendors who build their own call the result proprietary and disclose nothing. This vendor built its own and documented it further than any buyer here documents a supplier: architecture, corpus composition, token counts, the proprietary and public split, and the compute.

Building permits silence but plainly does not require it, and the firm with the most commercially valuable corpus in the market chose to publish. Held off A because no document maps which model powers which shipped feature, the base model paper is from 2023 and the newest agentic interface arrived in 2026, so a buyer cannot tell what is answering a given question today.

Meets the system of record test in several places at once rather than through the research platform. The company operates order and trade management systems that are the book of record for buy side and sell side desks, communications archiving used for supervisory retention, market data feeds delivered into customer infrastructure, and the security identifier scheme much of the market references. The Terminal is also the working environment its users spend the day inside, which is the position the generative features are being built into.

The vendor publishes no hosting model, region, residency commitment or tenancy position in the material reviewed. A claim exists in third party coverage that the financial language model runs only on the firm's own infrastructure with no external interface, which if stated by the vendor would be a meaningful containment disclosure, but it was not located on the vendor's own pages and earns nothing here. Queued check.

An unusual C worth distinguishing from an ordinary one. A single stable annual per seat figure of roughly 32,000 dollars is reported consistently across independent sources and is close to public knowledge in this market, so a buyer is not in the dark.

It is still graded C because the vendor publishes no price on its own pages, the standing rule in this index is never to grade commercial from an aggregator, and nothing states whether the generative features are included in the seat or charged separately, which is the question a reader of this record would actually be asking.

More than 325,000 Terminal subscribers worldwide is among the largest published user bases in the index, spanning the buy side, the sell side, corporate treasury and finance functions, government and public sector users, and media. Coverage is global rather than regional and the underlying content estate reaches over 200 million company documents. The generative features sit inside that same platform rather than being sold to a narrower segment.