Legal AI vendors are developing proprietary inference models to reduce cloud compute bills and minimize dependency on third-party platform APIs.
Legal AI vendors are increasingly developing their own artificial intelligence models to decrease dependence on large, external platforms like OpenAI and Anthropic. Companies such as Harvey and Thomson Reuters are moving toward vertically integrated AI stacks, a strategy anticipated to lower costs associated with model inference. Harvey recently announced Tenet, a custom model built specifically for legal work, while Thomson Reuters outlined plans to roll out its own model, called Thomson, in late July. Both initiatives rely on open-source foundation technology that can be downloaded and customized, marking a significant departure from the past two years, when many legal AI products functioned primarily as “model routers” that switched between OpenAI, Anthropic, and Google depending on the task.
These moves reflect a clear operational reality: legal AI vendors are trying to stop paying frontier-model tolls on every query and start owning more of the stack themselves. For enterprise legal operations leaders, this shift represents less a branding exercise about whose model is “best.” It’s a contract and governance change. When a vendor operates its own model, variables such as pricing mechanics, security ownership, and data portability all shift accordingly.
The most direct advantage for vendors is cost control. Industry advisers cited by Bloomberg Law expect that directing more queries to proprietary models rather than paying third-party providers for inference will improve vendor profitability. That matters to buyers because vendor margin pressure has a way of showing up later as usage limits, feature gating, or price resets at renewal.
However, the operational trade-off involves a shift in responsibility. Brenda Leong, director of the ZwillGen AI Division, noted that when a company runs its own model, it also takes on security and other obligations that were previously pushed down to the frontier lab. Maintenance and updates don’t disappear, they get reassigned. Consequently, legal AI procurement is turning into a stack decision: who pays for inference, who patches the model, and who proves it’s safe to use on sensitive work.
In practice, legal departments should read “custom model” as a signal to scrutinize the vendor’s model lifecycle: red-teaming cadence, evaluation sets for hallucination and citation behavior, and how quickly the vendor can roll fixes without breaking workflows. Those are procurement questions, not demo questions.
While firms like Harvey and Thomson Reuters work to minimize reliance on external frontier labs, OpenAI is pursuing a contrasting strategy aimed at closer alignment with legal buyers. According to Business Insider, OpenAI hired Jason Boehmig in June, an executive best known for building Ironclad into a major contract lifecycle management player. This appointment signals OpenAI’s intent to develop legal-specific distribution and product capabilities beyond simple API sales. The outlet also reported that OpenAI has been recruiting for roles tied to legal product development, though it declined to outline specifics of what it will build. Separately, Bloomberg Law reported that OpenAI is collaborating with Willkie Farr & Gallagher to deploy ChatGPT Enterprise to every attorney and integrate OpenAI models deeper into the firm’s internal tools. A firmwide rollout differs operationally from a pilot, as it forces decisions on identity, logging, matter segmentation, and acceptable use to be resolved at scale.
Historically, a vendor’s ability to switch among models was framed as resilience. Now it can also be a moving target. Legal tech companies have increasingly routed workloads across OpenAI, Anthropic, and Google, but the emerging trend of in-house model development suggests that the underlying architecture powering these products may change more frequently—and more quietly—than most legal departments have traditionally governed. This reality lands directly in procurement language. If your contract assumes a named subprocessor (for example, a specific frontier lab), and the vendor swaps to an in-house model, the risk profile changes even if the UI stays the same. Budgeting is equally affected, as vendor-hosted models can alter cost curves for high-volume users, particularly in contract review, discovery-adjacent analysis, and large-document summarization where token consumption is highly predictable.
Industry advisers note that average attorneys may not notice a change unless model quality shifts. That’s a useful reminder for operators: user satisfaction scores might stay flat while the underlying cost, security posture, and auditability changes materially. If the demo looks identical after a model swap, that’s when governance has to be strongest, because risk can move while workflows don’t.