Thomson Reuters says it spent $40 million on talent and computing power to build Thomson, a proprietary large language model the company announced Monday after starting with an open-source foundation.

The company is presenting that approach as a less expensive route to a specialized model. Thomson Reuters said frontier AI labs have typically spent billions of dollars on compute and years building infrastructure. It did not name those companies or compare the $40 million investment with the price of an AI content-licensing deal.

The model draws on decades of material from Westlaw, Practical Law, Checkpoint and Reuters. Thomson Reuters says hundreds of subject-matter experts were involved, from setting training objectives through final evaluations.

“For years, the AI industry has treated scale as the answer: bigger models, more compute, more money,” CTO Joel Hron said in the release. “Thomson shows there is another path.”

So far, the company says, Thomson has been trained on less than 10% of Thomson Reuters content.

CEO Steve Hasker said early internal evaluations put the model “on par with the latest frontier models across a range of tasks.” Thomson Reuters has also given the model to outside academics for testing.

Jonathan H. Choi of Washington University School of Law tested Thomson against ChatGPT and Claude on corporate tax questions. He said all three answered correctly, but preferred Thomson’s responses, especially its links to treatises. Samuel Dahan of Queen’s Conflict Analytics Lab said its citation quality was “generally competitive with leading frontier models” on Canadian employment-law questions.

A smaller open-weight version is also going up on Hugging Face for academic and non-commercial use.

Thomson’s first deployment is narrow. Tabular Analysis inside CoCounsel Legal, a tool for structured document review. CoCounsel will remain multi-model, using Thomson where it performs best and other providers’ models elsewhere.

The part likely to interest publishers is Thomson Reuters’ explanation for why specialization matters. The company says its early results challenge the idea that a general-purpose model can reach expert-level performance simply by getting access to the right content. Its claim is narrower: proprietary training and human expertise produced gains that content access alone did not.

That is not the same as saying publishers should stop licensing their archives or build models themselves. But it adds another approach to an industry conversation around negotiated content deals and technical countermeasures against scrapers: what can a company do when it controls both the content and the model?

Thomson Reuters is not presenting the project as a template for every publisher. The announcement supports the $40 million investment and the involvement of hundreds of subject-matter experts, but it does not say they were all in-house or identify which business units funded the work.

Reuters has separately been experimenting with AI in newsroom workflows. Editor-in-chief Alessandra Galloni has described AI as a way to shift routine production work away from reporters so they can spend more time gathering news in the field.

Thomson Reuters says it plans to expand the model across its legal and tax portfolio, with broader “sovereign AI options” to follow. It does not say those options will specifically include sovereign hosting.

The next test is whether Thomson’s early performance claims hold up as more independent evaluators get access and how much further the model can improve when less than a tenth of the company’s content has been used so far.

The post Thomson Reuters says it spent $40 million to build its own frontier AI model appeared first on The Media Copilot.

Meta is building a new AI model it plans to sell access to, abandoning the open-source strategy that Mark Zuckerberg championed for years.

The model, codenamed Avocado, is expected to launch next spring as a “closed” system that Meta can monetize directly, according to Bloomberg’s Kurt Wagner and Riley Griffin. This would align Meta with rivals OpenAI and Google, which charge for access to their most capable models.

The shift comes after Llama 4, Meta’s open-source model released earlier this year, disappointed Zuckerberg and Silicon Valley. He sidelined some team members and recruited top AI researchers with multiyear pay packages worth hundreds of millions of dollars.

New Chief AI Officer Alexandr Wang, who joined through Meta’s $14.3 billion acquisition of Scale AI, advocates for closed models. He now leads a group called TBD Lab that reports directly to Zuckerberg.

The team is training Avocado using several third-party models, including Qwen from Chinese tech giant Alibaba. That’s a notable turn for Zuckerberg, who warned on Joe Rogan’s podcast in January that Chinese models could be shaped by state censorship.

Yann LeCun, known as one of AI’s “godfathers” and a major open-source proponent, left Meta recently after years leading its long-term AI research. Some employees had been encouraged to keep him out of the spotlight, Bloomberg reported, because Meta no longer saw him as representative of its AI strategy.

Why it matters for newsrooms: Media organizations evaluating AI tools now face a landscape where even Meta, once the loudest voice for open AI development, is moving toward paid, proprietary systems. That could mean fewer free options for cash-strapped newsrooms and greater dependence on subscription-based AI services from a handful of tech giants.

The post Meta abandons open-source AI strategy in pursuit of paid model appeared first on The Media Copilot.