Thomson is a large language model built by Thomson Reuters for legal and tax work. The company published first benchmark results on July 31, 2026 in a post by Chief Technology Officer Joel Hron and Head of AI Research Jonathan Schwarz (Thomson Reuters Built Its Own AI Model That Now Ranks Among the World's Best), and deployed the model on August 24, 2026 as the default model behind the Tabular Analysis document-review capability of CoCounsel Legal (Source: siliconangle.com).
The model is notable less for its capability level than for who built it: an incumbent legal-information provider training its own domain model rather than layering products over a frontier lab's API, while continuing to build other parts of the same product line on Anthropic's Claude Agent SDK.
All published performance figures are the developer's own. No technical report, system card, parameter count, or independent evaluation had been published as of August 25, 2026.
Naming
The company's prose calls the model "Thomson"; its benchmark table labels the tested system "Thomson-1-Large." Whether these name the same system, or whether Thomson-1-Large is one variant of a family, is not stated in the published material.
Training and architecture
Thomson Reuters states that the model "starts from a strong open-source foundation," without identifying it, and is then trained with mid-training and post-training techniques on content from Westlaw, Practical Law, Checkpoint and Reuters. Training was shaped by "hundreds of subject matter experts" who evaluated outputs and identified failure modes. The company says customer data is never used to train the model, and that fewer than 10% of its content holdings had been used in training as of July 2026 (Thomson Reuters Built Its Own AI Model That Now Ranks Among the World's Best).
The company traces the capability to its 2024 acquisition of Safe Sign Technologies, an AI research company (Thomson Reuters Built Its Own AI Model That Now Ranks Among the World's Best).
Later reporting puts the project's cost at about $40 million over two years in people and computing, with efficiencies bringing the final training run to roughly $450,000. Senior research scientist Andrew Bean said internal tests showed the model broadly competitive with leading models when all had web access only, and roughly equal or slightly better when connected to Thomson Reuters content; a technical report was said to be expected later (Source: siliconangle.com).
Capabilities and benchmarks
The July 2026 disclosure published seven benchmark rows for Thomson-1-Large against Gemini 3.1 Pro, Claude Opus 4.8, and GPT-5.5. The comparison is not like-for-like: the company states that Thomson-1-Large used test-time scaling, Gemini 3.1 Pro and Opus 4.8 used reasoning mode, and GPT-5.5 used non-reasoning mode.
| 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 |
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 benchmarks (Thomson Reuters Built Its Own AI Model That Now Ranks Among the World's Best).
On the seven published rows Thomson-1-Large takes the best score on three — PrBench Legal Hard, Instruction Following, and Long Context — and trails on the two general-capability rows where the gap is widest, Reasoning and Coding. The company's own summary claim, that the model performed "ahead of GPT-5.5, Claude Sonnet 5, and Gemini 3.1 Pro," is broader than the table supports: Claude Sonnet 5 is absent from it, and Thomson-1-Large leads each of GPT-5.5 and Gemini 3.1 Pro on four rows of seven.
A second evaluation covered 53 legal research queries written by internal subject-matter experts, with models connected through an in-house agentic harness to Westlaw and Practical Law and to the Brave search engine for web search. Completeness was scored against expert-written rubrics and factuality by extracting each report's claims and checking the cited sources, both judged by language models calibrated against expert scoring. The results were published only as a chart, and no values are recoverable from it (Thomson Reuters Built Its Own AI Model That Now Ranks Among the World's Best).
Availability
Thomson is not offered as a standalone API. It reached users on August 24, 2026 inside Tabular Analysis in CoCounsel Legal, where it is the default model and administrators retain the ability to select others. The company has said it will integrate the model across its legal and tax portfolio over the following year, and plans to release a smaller open-weight version on Hugging Face under a noncommercial academic licence (Source: siliconangle.com; Thomson Reuters Built Its Own AI Model That Now Ranks Among the World's Best).
Hron, described in the August account as global head of artificial intelligence and TR Labs, said Thomson "needs to set the frontier of intelligence for legal," which he distinguished from the job of the frontier labs (Source: siliconangle.com).
Safety and evaluations
The developer states that it conducts "training and evaluations consistent with Thomson Reuters values" and stress-tests the model "through human and automated red-teaming," without publishing results, a methodology, or a threshold (Thomson Reuters Built Its Own AI Model That Now Ranks Among the World's Best). No system card, model card, or safety framework accompanied either the benchmark disclosure or the deployment.
Related models
Thomson is a domain-specialised model competing on legal tasks with general-purpose frontier models — Gemini 3.1 Pro, Claude Opus 4.8, and GPT-5.5 in its own comparison — rather than with them on general capability, where its published Reasoning and Coding scores trail. Its nearest structural comparison is not another model but the strategy of Harvey and Legora, which build legal products on frontier-lab APIs; Thomson Reuters continues to do both, running CoCounsel Legal on Anthropic's Claude Agent SDK while Tabular Analysis runs on Thomson.
Relationships
- developer: Thomson Reuters
- deploys-in: Legal Services — AI Deployment — sold into law firms, corporate legal departments and government legal offices through CoCounsel.
- instance-of: Professional Services AI Adoption — a domain model built by an information incumbent.
- depends-on: AI Benchmarks and Evaluation — the model's public claim rests entirely on self-reported, non-like-for-like benchmarks.
- related: Harvey, Legora — legal-AI vendors taking the build-on-frontier-models approach.
- related: Anthropic — supplies the Claude Agent SDK underlying CoCounsel Legal alongside Thomson.
Sources
- Thomson Reuters Built Its Own AI Model That Now Ranks Among the World's Best — Hron and Schwarz, Thomson Reuters, July 31, 2026.
- SiliconANGLE, "Thomson Reuters launches proprietary AI model for legal work," August 24, 2026 (Source: siliconangle.com).
- Thomson Reuters press release, next generation of CoCounsel Legal, August 20, 2026 (Source: thomsonreuters.com).