Thomson Reuters has launched a custom AI model, Thomson, stepping up its push to establish itself as an artificial-intelligence developer while shielding itself from the industry’s mounting costs.

The stakes: Monday’s launch could be a key test of whether domain-specific large language models (LLMs) are ready to compete with generalized giants, as more companies experiment with custom AI to better control their data and their budgets. 

The Toronto-based firm, controlled by one of Canada’s richest families, has found its fortunes increasingly tied to investor sentiment around AI. In February, Thomson Reuters’ stock plummeted after Anthropic launched a legal tech tool, stoking fears that general-purpose AI would replace niche legal software. Days later, Thomson Reuters sought to reassure investors by working closely with Anthropic’s Claude on new features. 

A new concern has since emerged across the software sector: frontier AI models like Claude, ChatGPT or Gemini are too costly. Companies like Uber are burning through AI budgets, prompting providers like Google to slash subscription prices.

The headline numbers: On an earnings call this month, Thomson Reuters CEO Steve Hasker said the company spent US$40 million training its Thomson model, a fraction of what frontier labs spend. The stock still fell nearly 10 per cent on the afternoon the report was released. 

Chief technology officer Joel Hron said in a press briefing last week that the final training run for Thomson will cost roughly US$450,000.

Thomson Reuters may be spending far less than Anthropic or OpenAI, with its Thomson model potentially giving the company more bargaining power with suppliers. But its AI investments will also cut into profit margins that are already tighter than investors were expecting, CIBC analysts wrote in a research note. 

The Thomson model is based loosely on Alibaba’s open source model Qwen, but could be run on other models if better options come to market. Executives said a secondary model screens and realigns the open source technology for Thomson Reuters’ safety, ethics and political neutrality standards. The company is training Thomson to focus on areas like journalism, law and taxes, rather than investing in fancier features like coding.

The news behind the legalese: Starting this week, lawyers, software developers and investors can put the new LLM through its paces. The Thomson model will be released on Hugging Face, a major hub for software developers to test out new AI models, which made headlines when OpenAI models targeted it in a hack last month. 

Law firms, meanwhile, will get their first look through the CoCounsel chatbot, which will switch over to using Thomson for some advanced features this week.

The big picture: Thomson Reuters isn’t alone in betting on domain-specific LLMs. Early attempts like BloombergGPT struggled against off-the-shelf models, but AI companies like Mistral have become more bullish on LLM customization, arguing domain-specific models could offer greater data control and bigger improvements in AI performance. 

Legal tech entrepreneur Gordon Cassie, founder of Eno PDF, has been working ways to improve how AI chatbots deliver citations. In an interview before Thomson Reuters announced its LLM launch, he noted that the increased adoption of open-weight AI models has encouraged more companies to experiment with LLM customization.

Cassie acknowledged that some companies may be promoting custom AI to protect their software businesses, but said tailored models can also address problems with off-the-shelf AI, including hallucinations in court filings. 

Large frontier labs have captured the majority of the AI boom’s value so far. Cassie said it remains an “open question” how they can maintain that lead, as falling prices and competition from open source or customized alternatives put pressure on their businesses. 

“There are now big, increasing questions of who is actually, business wise, going to win in all of this,” he said.