Thomson Reuters has invested approximately $40 million to develop Thomson-1, its first proprietary large language model (LLM). This move signifies a strategic shift for the information services giant, aiming to reduce its dependence on outside artificial intelligence providers for core technologies. The company’s chief technology officer described the initiative as building long-term equity in its AI capabilities.
The launch of Thomson-1 represents a calculated step toward owning the technology that underpins its flagship products. While its marquee AI assistant for lawyers, CoCounsel, still primarily uses Anthropic’s Claude, Thomson-1 is designed for specific high-volume tasks where deep domain knowledge offers a measurable advantage. This approach contrasts with the broader trend of relying solely on general-purpose AI models.
Developing a Proprietary AI Model
Thomson-1 is built upon an open-source foundation from Alibaba’s Qwen model, which a joint team with Imperial College London adapted. This adaptation process, described by Joel Hron, the company’s CTO, involved months of work. The focus was on incorporating ethical safeguards, de-biasing mechanisms, and safety protocols into the model.
The LLM was trained using less than 10% of Thomson Reuters’ extensive proprietary corpus. This content includes decades of authoritative legal, tax, and news information. Hundreds of subject-matter experts rigorously reviewed the model’s outputs. They worked to identify failures and refine the system, prioritizing accuracy over simply generating pleasing responses.
Performance and Strategic Integration
Early benchmarks released by Thomson Reuters in late July indicate that Thomson-1 performs competitively with advanced frontier systems. The model matched or exceeded Claude Opus 4.8 in certain legal reasoning tasks. It also outperformed GPT-5.5, Claude Sonnet 5, and Gemini 3.1 Pro across a range of evaluations. These tests covered instruction following, long context processing, coding, and professional workflows.
The company noted that these comparisons were not always “apples to apples,” as Thomson-1 benefited from test-time scaling and internal retrieval tools linked to products like Westlaw and Practical Law. In some evaluations, competing models searched the open web. Despite these caveats, the results impressed academics, with some preferring Thomson’s responses for their citation quality, even when other models provided correct answers. This development highlights the increasing integration of AI adoption across law firms and legal teams.
Building Internal Equity
Joel Hron emphasized the strategic value of this investment. He likened renting a house to relying on external AI providers, stating, “Renting a house, you still have a roof over your head, and somebody’s taking care of it, and it’s great. But you’re not building any equity that compounds into something valuable for you long term.” The $40 million expenditure on compute resources, talent, and specialized training aims to build that internal equity.
This investment reflects a broader trend within the legal sector to understand and leverage new technologies. Many legal professionals are now focusing on hands-on AI learning to adapt to industry changes. The company’s strategy positions Thomson Reuters to have greater control over its AI development. It also allows for deeper customization based on its unique content and specific legal industry needs. This move also highlights how the legal industry navigates innovation while ensuring reliability.
The initial real-world application of Thomson-1 will be within CoCounsel Legal, suggesting a phased integration into Thomson Reuters’ broader product ecosystem. This strategic investment in proprietary AI could redefine how major legal information providers build and deliver their services, fostering greater independence and specialized capability. The company can adapt the model’s foundation as needed, as Hron noted, “There’s nothing that necessarily ties us to Qwen.”
