Thomson Reuters builds own AI for $40M, skips OpenAI

Thomson Reuters' decision to build its own model rests on a specific economic logic: the real capital is not the $40 million or even the $450,000 final training run, but decades of content from Westlaw, Practical Law, Checkpoint, and Reuters, plus the working hours of hundreds of domain experts. The company's own benchmarks show why this matters. Thomson trails on reasoning and coding, but leads on instruction following and the PrBench Legal benchmark, and only surpasses GPT 5.4 when it can tap into its own content. This suggests the data is doing as much work as the specialized training.

The company's framing is careful about what it claims. Less than 10 percent of available content has gone into training so far, which implies room for improvement. CTO Joel Hron notes they have changed the open-source starting point close to half a dozen times already, and research chief Jonathan Schwartz emphasizes the "model factory" over any individual model. The comparison with frontier models is also lopsided: Thomson competes with test-time scaling, while GPT-5.5 runs without reasoning mode, and the company did not test newer models.

For companies with a similar profile, the case suggests the open-source community trails frontier labs by only months, and $40 million can be enough for a competitive specialized model. The deeper point, as Hron puts it, is that the next competitive advantage in AI will come from knowing how to orchestrate it and knowing which intelligence is important enough to own. Whether Thomson's razor-thin lead holds will depend on whether it moves to a stronger variant like Qwen3.8 and pushes past the 10 percent content mark.