Thomson Reuters Built Its Own AI Model That Now Ranks Among the World's Best is a post published on the Thomson Reuters website on July 31, 2026 (modified August 3) and credited to Joel Hron, Chief Technology Officer, and Jonathan Schwarz, Head of AI Research (Source: thomsonreuters.com). It is the first public disclosure of benchmark results for Thomson, a large language model Thomson Reuters trained itself for legal and tax work, and it sets out the company's argument for building a domain model rather than buying general-purpose capability.
The post is a vendor self-report. Its scores are the company's own, had no independent validation at publication, and — by the company's own note — were produced under inference settings that differ between the compared models.
Argument
The post opens by asserting that "the most capable AI models no longer come only from frontier AI labs." It traces the model to the company's 2024 acquisition of Safe Sign Technologies, an AI research company, and frames that purchase as a bet against the prevailing view that "access to increasingly powerful general-purpose models would be enough." The company's counter-position is that professional AI "would require models built specifically for the domains, standards, and consequences of professional work," and that its content and expertise "could be best expressed in a model that we ourselves crafted."
Thomson is described as starting "from a strong open-source foundation" — the post does not identify which — so that it performs general-purpose work as well as frontier models, then trained with mid-training and post-training techniques on decades of content from Westlaw, Practical Law, Checkpoint and Reuters. Training was shaped by "hundreds of subject matter experts" who evaluated outputs, identified failure modes, and validated that the model "reasons the way legal professionals actually work." The company states that customer data is never used to train the model.
The post markets the result under a company trademark, Fiduciary-Grade AI™, defined as "AI designed for professionals with duties of care and accountability, where almost right is not good enough."
Reported benchmarks
The post publishes a head-to-head table for the variant it labels Thomson-1-Large against Gemini 3.1 Pro, Claude Opus 4.8, and GPT-5.5. The values below are the published ones; the best score in each row was highlighted in the original.
| Benchmark | Thomson-1-Large | Gemini 3.1 Pro | Opus 4.8 | GPT-5.5 |
|---|---|---|---|---|
| Legal — Stanford LegalBench | 0.823 | 0.843 | 0.818 | 0.832 |
| Legal — PrBench Legal Hard | 0.352 | 0.293 | 0.315 | 0.333 |
| Legal — Harvey Legal Agent Benchmark | 0.857 | 0.555 | 0.869 | 0.781 |
| General — Instruction Following | 0.914 | 0.848 | 0.861 | 0.885 |
| General — Reasoning | 0.684 | 0.748 | 0.737 | 0.589 |
| General — Coding | 0.399 | 0.500 | 0.598 | 0.414 |
| General — Long Context | 0.753 | 0.750 | 0.741 | 0.703 |
Four rows are composites: Instruction Following averages IFEval and FollowBench; Reasoning averages GPQA Diamond, HLE and MMLU-Pro; Coding averages SWE-Bench Pro and Terminal-Bench 2.1; Long Context averages Infinity Bench with internal Thomson Reuters benchmarks.
The company states in a note to the table that Thomson-1-Large used test-time scaling, that Gemini 3.1 Pro and Opus 4.8 used reasoning mode, and that GPT-5.5 used non-reasoning mode. The comparison is therefore not like-for-like across inference settings, and any use of the table carries that qualification.
The post also lists the evaluation categories it says it covered — legal, tax, accounting, journalism, safety, reasoning, coding, math, multilingualism, agentic tasks, long context, and instruction following — and describes internal work beyond the published table: agentic use cases, "tens of thousands of real-world queries written by experts," end-to-end deep research training with human-calibrated judges, data-centric mid-training, and human and automated red-teaming.
Divergence between the lead and the table
The opening paragraph claims Thomson performed "competitively with the strongest frontier models on the market, including Claude Opus 4.8, and ahead of GPT-5.5, Claude Sonnet 5, and Gemini 3.1 Pro." Claude Sonnet 5 does not appear in the published table, and on the seven published rows Thomson-1-Large leads GPT-5.5 on four and Gemini 3.1 Pro on four. The summary claim is broader than what the table shows.
Retrieval-augmented evaluation
A second evaluation, published only as a chart with no recoverable values, covered 53 legal research queries written by internal subject-matter experts to represent real-world questions. Models were connected through an in-house agentic harness to Westlaw and Practical Law for company content and to the Brave search engine for web search. Completeness was scored against expert-written rubrics listing every element required for a good answer; factuality was scored by extracting the claims in each report and checking whether the cited sources supported them. Both were scored by language models acting as judges, calibrated against subject-matter-expert scoring. The company's stated conclusion is that Thomson's access to Westlaw, Practical Law and Reuters "ensures both superior completeness and factuality" against frontier models given unrestricted web access; no numbers accompany that claim in the retrievable text.
Scale and deployment as stated
The post says Thomson is "a fraction of their size and cost to train and operate" relative to the general-purpose models compared, without giving a parameter count, and that "less than 10% of Thomson Reuters content has been used in its training so far."
It announces the first integration as Tabular Analysis inside CoCounsel Legal, launching "in August," where Thomson would become the default model. The company's stated reason for that choice is that Tabular Analysis performs "high-volume, structured document review against a clear, measurable accuracy standard," making the advantage of a purpose-built model measurable. Integration "across the Thomson Reuters product portfolio in legal and tax" is described as planned over the following year.
Provenance and limits
Published on the company's own domain under the "AI and product innovation" topic. Page metadata gives the authors as Joel Hron and Jonathan Schwarz and a publication date of July 31, 2026, modified August 3. The Thomson Reuters press release of August 20, 2026 announcing the next generation of CoCounsel Legal links to this post as its reference for the Thomson model.
The captured text is complete. The benchmark table was published as an image and its values are transcribed from the image alt text; the completeness/factuality chart carries no recoverable values, so only its methodology is available. No parameter count, training-compute figure, open-source base model, or technical report accompanied the post. Deployment date, project cost, and the planned open-weight release reach the record through separate later reporting rather than through this source.
Relationships
- supports: Thomson (Thomson Reuters legal-domain model) — the primary published record of the model's reported benchmarks and training approach.
- instance-of: Professional Services AI Adoption — an information incumbent training a domain model rather than relying on general-purpose ones.
- related: Legal Services — AI Deployment — the sector the model is built for.
- related: Harvey, Legora — legal-AI vendors building on general-purpose frontier models, the approach this post argues against.
- related: AI Benchmarks and Evaluation — the post's self-reported, non-like-for-like comparison is an instance of the vendor-benchmark problem.