Thomson Reuters launches $40M legal LLM inside CoCounsel

Why it matters

Thomson shows how data-rich incumbents can use open weights to own specialized AI infrastructure without abandoning frontier-model suppliers.

Thomson Reuters launched Thomson on August 24th, its first proprietary large language model and a $40 million bid to own more of the AI infrastructure behind its legal and tax products.

The model grew out of work led by Alexander Kardos-Nyheim and Jonathan Schwarz, who built legal language models at Safe Sign Technologies before Thomson Reuters acquired the UK business in August 2024. Safe Sign was founded in 2022 and became Thomson Reuters' first pre-revenue acquisition, according to Kardos-Nyheim. Its 15-person operation later became the core of a 50-person Foundational Research group inside Thomson Reuters.

Thomson Reuters says the $40 million figure covers talent and compute, so it should not be read as a pure model-training bill. Thomson starts with the open-weight Snowdon foundation model developed at Imperial College London, then applies mid-training and post-training using material from Westlaw, Practical Law, Checkpoint and Reuters. Hundreds of subject-matter experts helped set training objectives and evaluate the results, according to Thomson Reuters.

That approach follows the argument Kardos-Nyheim and Schwarz made at Safe Sign: general model capability would become widely available, while reliable performance in law and other regulated fields would depend on specialized training, evaluation and proprietary material. In an account of how Thomson was built, Kardos-Nyheim wrote that customers were also questioning their dependence on outside model architectures, pricing and product roadmaps.

Thomson Reuters says Thomson has so far trained on less than 10% of its proprietary content. Thomson Reuters claims the model performs competitively with Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro across legal and general-purpose evaluations, while costing less to train and operate than comparable frontier models.

The benchmark table Thomson Reuters published in July gives a narrower picture. Thomson posted the best result in three of seven reported categories: PrBench Legal Hard, instruction following and long-context performance. It placed behind another model on Stanford LegalBench, the Harvey Legal Agent Benchmark, general reasoning and coding.

The comparison also used different inference settings. Thomson ran with test-time scaling, Gemini 3.1 Pro and Claude Opus 4.8 used reasoning modes, and GPT-5.5 ran without a reasoning mode. Thomson Reuters' long-context score combined a public benchmark with internal evaluations. Those qualifications make the results useful evidence of a specialized model's potential, rather than an independent verdict that Thomson has surpassed the general-purpose frontier.

Thomson Reuters has begun giving legal and AI academics access for direct evaluation. Jonathan H. Choi of Washington University School of Law tested Thomson alongside ChatGPT and Claude on corporate tax questions and reported that all three answered correctly, with Thomson providing the responses he preferred. Samuel Dahan, who directs legal AI research groups at Queen's University and Cornell, found its citation quality generally competitive in a Canadian employment-law test. Thomson Reuters also said it is releasing a smaller open-weight version for academic and non-commercial research.

Customers will first encounter Thomson through Tabular Analysis in CoCounsel Legal. The feature can review as many as 10,000 documents against 100 questions, returning answers linked to their sources. Thomson Reuters plans to make Thomson available there in an upcoming release before extending versions of the model across its legal and tax portfolio.

Thomson is one layer in a multi-model system. Thomson Reuters says CoCounsel will continue routing work to outside models when they perform better, and the next generation of CoCounsel Legal uses Anthropic's Claude Agent SDK to plan and execute workflows. Thomson is not currently sold as a standalone model.

That structure explains the business case. Thomson Reuters can direct high-volume, domain-specific work to infrastructure it controls while retaining access to the strongest general-purpose systems. Ownership gives Thomson Reuters another lever over inference costs, data governance and supplier dependence without forcing CoCounsel to standardize on a single model.

The model also gives the Safe Sign founders a production test for the thesis that led Thomson Reuters to buy their research operation before it had revenue. The archive, experts and distribution now come from Thomson Reuters. Kardos-Nyheim and Schwarz supplied the model-training operation intended to turn those assets into a system Thomson Reuters can own.