Thomson Reuters has begun moving some of its legal AI workloads from Anthropic’s Claude to Thomson-1, an in-house model built using Chinese open-source technology. The change is expected to lower costs, while also raising questions about the security and governance of Chinese AI systems in a still-evolving industry.

According to Business Insider, Thomson-1 is set to handle the document-review tasks previously performed by Claude. Legal professionals, accountants, and companies using CoCounsel and similar products are unlikely to see major changes in day-to-day use. CoCounsel joined the Thomson Reuters portfolio through its $650 million acquisition of legal-AI startup Casetext in 2023. The more notable development is that the company behind the Westlaw legal research service — a fixture of Western legal practice — is relying on a Chinese model that has been extensively modified in-house.

Cheaper than Claude, by design

Cost was a major factor in the decision. Thomson Reuters CTO Joel Hron told Business Insider that the high prices charged by labs such as Anthropic and OpenAI encouraged the company to develop its own model. By building its own system, Thomson Reuters can use its intellectual property rather than continue paying third-party service fees.

Hron said Thomson-1 will not fully replace Anthropic. Thomson Reuters expanded its partnership with Anthropic in May, and CoCounsel still relies heavily on Claude. Instead, the company plans to move gradually and use Thomson-1 only where doing so benefits its own expertise. For now, the system is intended for “high-volume, structured document review.”

What “realigning” Qwen means

Thomson-1 is based on Snowdon, which was created by “realigning” an open-source Qwen model from Alibaba. Unlike Claude, which is accessed as a closed system through Anthropic, Qwen’s open weights can be downloaded and modified — and, once downloaded, run on a company’s own infrastructure, so the documents being reviewed need not pass through an outside vendor’s API.

A Thomson Reuters–Imperial College London team spent months adapting Qwen into Snowdon. Hron described the result as “ethically and politically de-biased and safe to use.” Cryptopolitan has previously reported that Western companies are increasingly adapting foreign open-weight models to gain more control over their AI stacks. Hron also said, “there’s nothing that necessarily ties us to Qwen.”

Thomson Reuters says Thomson-1 performs competitively with Claude Opus 4.8 and ahead of GPT-5.5, Claude Sonnet 5, and Gemini 3.1 Pro across its broader evaluation suite. Those results are company-reported, and independent academic benchmarking is still underway. The published scores suggest a mixed picture rather than a clean sweep.

Concerns about Chinese open-source models

The cost savings come with political and security concerns. Anthropic has accused Chinese labs of illegally “distilling” its model outputs and has urged Washington to impose restrictions; some US state governments have already barred Chinese AI apps such as DeepSeek from official devices over data-security concerns. Senator Tom Cotton has also raised security concerns about US companies using Chinese open-source models.

Stanford researchers have raised a different issue. HAI’s James Landay said downloadable model weights do not make a system fully auditable because users still cannot see the training data or understand every behavior. He described this as “open distribution” rather than open source. That makes it difficult for customers to independently verify claims that bias has been removed — a question with particular weight in legal work, where document review routinely involves privileged and confidential client material.

A gap measured in months

Companies are willing to consider the trade-off because Chinese models have advanced quickly. Stanford’s 2026 AI Index said the US-China performance gap had “effectively closed.” On Arena’s Text leaderboard as of August 21, 2026, the leading US model scored 1,508, compared with 1,489 for the top Chinese model, a gap of just 1.3%. Alibaba’s Qwen3.8-Max scored 1,481.

When capability differs by only a few points while costs remain much higher, document review becomes a natural testing ground for cheaper alternatives.