Key Takeaways
- Thomson Reuters launched a $40 million proprietary LLM and a separate platform built on Anthropic’s Claude.
- The company employs a hybrid AI model strategy, balancing cost-optimized in-house inference with frontier model licensing.
- GenAI-enabled products now represent 32% of Thomson Reuters’ annualized contract value as of Q2 2026. Thomson Reuters is running two AI bets at once. On August 24, 2026, it launched Thomson, a proprietary LLM built on Alibaba’s Qwen and backed by a $40 million investment. Four days earlier, it had already pushed the next-generation CoCounsel Legal to general availability, built directly on Anthropic’s Claude Agent SDK. The same company, the same week, building its own model and deepening its reliance on a frontier provider.
A $40 Million Bet on Specialisation
The model adapts Alibaba’s open-source Qwen3.5, undergoing an intermediary process aimed at de-biasing the base model, both ethically and politically, before further domain adaptation.
During the Q2 2026 earnings call, the company reported that Thomson performs on par with leading frontier models on general tasks and stronger on legal-specific ones, at lower cost and latency. The goal is to embed Thomson Reuters’ proprietary content, Westlaw, Practical Law and decades of legal and tax data, directly into the model’s architecture, producing outputs that meet what the company calls “Fiduciary-Grade AI” standards: accurate, traceable and defensible in professional practice.
Anthropic’s Role in CoCounsel
For CoCounsel Legal Thomson Reuters went the other direction. The next-generation platform, generally available from August 20, 2026, is built on Anthropic’s Claude Agent SDK, giving it the multi-step planning and execution capabilities that define current agentic AI. The platform covers litigation workflows end to end: Westlaw research, document analysis, brief drafting via Westlaw Brief Builder, verification and matter management in a single connected environment. For complex, open-ended legal work, the reasoning depth of a frontier model still justifies its cost.
CoCounsel Legal launched in the US first, with the UK, Canada, and Australia scheduled to follow later in 2026.
The Cost Logic Behind the Hybrid
The build-vs-buy split is an economic calculation as much as a technical one. Licensing a frontier model for every query across Thomson Reuters’ enterprise customer base would compound quickly. For structured, repeatable tasks, document review, clause extraction, jurisdiction-specific research, a fine-tuned in-house model running at lower cost per token is the more defensible long-run position. The $40 million capital outlay for Thomson is, in that framing, an inference cost hedge.
For CoCounsel, the calculation runs the other way. Replicating Claude’s agentic SDK capabilities from scratch would require investment and development timelines that likely dwarf the licensing cost, at least at the current stage of the technology. As agentic frameworks mature and open-source alternatives close the gap, that calculus may shift, but for now, buying frontier capability and building domain-specific capacity in parallel is how Thomson Reuters is managing the trade-off. This approach also gives the company data sovereignty over its most sensitive workloads, routing them through Thomson rather than external APIs, while still delivering market-leading functionality where clients expect it.
Agentic Adoption: The Numbers
Raghu Ramanathan, President of Legal Professionals at Thomson Reuters, has described the shift as moving beyond AI that generates answers toward AI that completes legal work. The 2026 AI in Professional Services Report from the Thomson Reuters Institute puts some numbers to that. The report, drawing on responses from more than 1,500 professionals, found that only about 15% of organisations currently use agentic AI, but more than half are planning or actively considering adoption. A further finding: the vast majority of professionals surveyed expect agentic AI to be central to their workflows by 2030.
A further finding: three-quarters of professionals surveyed expect agentic AI to be central to their workflows by 2030. General AI adoption across professional services has also accelerated sharply, with 40% of professionals in the survey reporting organisational use, up from 22% the prior year.
For legal specifically, a significant majority of professionals already using generative AI apply it to legal research. The adoption curve is moving faster than ROI measurement: only 18% of professionals say their organisations actively track AI return on investment, while 40% are unaware whether it is tracked at all.
For more on Anthropic’s push into legal workflows and how competing agentic platforms are developing, our Enterprise AI coverage tracks these deployments as they scale. The ROI measurement problem Thomson Reuters is navigating is also visible in broader enterprise GenAI pilot failures where the gap between adoption speed and strategic measurement is the most consistent obstacle.
Open Source, Geopolitics and Qwen
Building on Alibaba’s Qwen3.5 introduces a dimension that goes beyond model performance. Open-source foundations cut development cost and allow full architectural control, but the provenance of the base model matters to enterprise customers, particularly those in regulated industries with data residency obligations or clients in Western government sectors. Thomson Reuters’ decision to route Qwen through Snowdon, with an explicit de-biasing step developed alongside Imperial College London, reads as a direct attempt to address that concern before customers raise it.
Open-source adaptation is a different risk profile from closed-source API licensing: more control, more responsibility, more exposure if something goes wrong. The intellectual property picture is also cleaner, no downstream usage restrictions from a proprietary model provider, and no dependency on a vendor’s API availability or pricing decisions. For a company that positions its AI output as fiduciary-grade, controlling the full stack from base model to fine-tuning is a defensible position. Whether the Snowdon de-biasing process holds up to external scrutiny is a question the market will eventually ask. Research on LLM ethical inconsistency suggests that de-biasing claims warrant independent verification rather than acceptance at face value.
What the Numbers Show
Thomson Reuters’ Q2 2026 earnings call, held August 5, 2026, reported that GenAI-enabled products accounted for 32% of annualized contract value, up from 30% the prior quarter. CEO Steve Hasker described revenue growth as stronger than expected, driven by demand for AI-enabled professional workflow tools. Two percentage points of ACV shift in a single quarter is a fast-moving revenue mix, and it suggests customers are actively buying the AI-integrated products, not just trialling them.
The 16% share price drop in February, triggered by Anthropic’s direct legal market move, remains the sharpest signal of how much market risk Thomson Reuters carries from its frontier model dependencies. The hybrid strategy is, in part, a structural answer to that risk: if any single external provider pivots, Thomson Reuters retains an in-house model capable of carrying core workloads. Whether Thomson can scale to cover more of the product portfolio over time, or whether CoCounsel’s reliance on Claude deepens further, will be the test of whether the $40 million investment becomes the foundation of something larger or remains a cost-management tool at the margins.
Originally published at https://autonainews.com/thomson-reuters-builds-its-own-40m-llm-while-expanding-claude-use/
Top comments (0)