Across a range of legal and general-purpose benchmarks, they wrote, Thomson performed competitively with Anthropic\u2019s Claude Opus 4.8 and ahead of OpenAI\u2019s GPT-5.5, Anthropic\u2019s Claude Sonnet 5, and Google\u2019s Gemini 3.1 Pro.<\/p>\n
\u201cThe most capable AI models no longer come only from frontier AI labs,\u201d Hron and Schwarz wrote. \u201cOne now comes from Thomson Reuters.\u201d<\/p>\n
I\u2019ll go into what the numbers actually show, how the tests were run, and what remains unverified, but first, some background on Thomson.<\/p>\n
What Thomson Is<\/strong><\/h3>\n\u201cLaunching later this summer, Thomson is the newest layer of the Thomson Reuters AI strategy, and a demonstration of what becomes possible when authoritative content, expert judgment, professional tools, and model development come together,\u201d Hron and Schwarz say in their post.<\/p>\n
According to the post, Thomson starts from an open-source foundation model and is then trained, using what they describe as state-of-the-art mid-training and post-training techniques, on decades of proprietary content from Westlaw, Practical Law, Checkpoint and Reuters.<\/p>\n
Hundreds of subject-matter experts evaluated outputs, identified failure modes, and validated the model\u2019s reasoning, they say, adding that customer data is never used in training.<\/p>\n
Notably, they say that less than 10% of Thomson Reuters\u2019 proprietary content has been used in training so far \u2013 suggesting there is plenty of room for Thomson\u2019s capabilities to further improve.<\/p>\n
The model\u2019s first deployment will come in August, when Thomson becomes the default model powering Tabular Analysis in CoCounsel Legal.<\/p>\n
Hron and Schwarz say they chose Tabular Analysis because that type of high-volume, structured document review offers a clear, measurable accuracy standard where a purpose-built model\u2019s advantage should be immediately visible. Over the next year, they write, Thomson Reuters will integrate the model across its legal and tax product portfolio.<\/p>\n
The Numbers<\/strong><\/h3>\n![\"\"](\"https:\/\/i0.wp.com\/www.lawnext.com\/wp-content\/uploads\/2026\/08\/TR8347338_02A_1080x1080-1-1024x1024.jpg?resize=1024%2C1024&ssl=1\" "\\"\\"")
<\/a><\/p>\nThe benchmark that drove the headline pits Thomson (technically called Thomson-1-Large) against Gemini 3.1 Pro, Claude Opus 4.8, and GPT-5.5 across three legal benchmarks and four general-domain composites. Thomson took the top score in three of the seven rows:<\/p>\n
Legal domain:<\/strong><\/p>\n\n- Stanford LegalBench:<\/strong> Thomson 0.823 \u2013 behind Gemini 3.1 Pro (0.843) and GPT-5.5 (0.832), ahead of Opus 4.8 (0.818).<\/li>\n
- PrBench Legal Hard:<\/strong> Thomson 0.352 \u2013 best in class, ahead of GPT-5.5 (0.333), Opus 4.8 (0.315), and Gemini (0.293).<\/li>\n
- Harvey Legal Agent Benchmark:<\/strong> Thomson 0.857 \u2013 a close second to Opus 4.8 (0.869), and well ahead of Gemini (0.555).<\/li>\n<\/ul>\n
General domain:<\/strong><\/p>\n\n- Instruction following<\/strong> (IFEval + FollowBench): Thomson 0.914 \u2013 best in class.<\/li>\n
- Reasoning<\/strong> (GPQA Diamond, HLE, MMLU-Pro): Thomson 0.684 \u2013 behind Gemini (0.748) and Opus (0.737).<\/li>\n
- Coding<\/strong> (SWE-Bench Pro, Terminal-Bench 2.1): Thomson 0.399 \u2013 last place, well behind Opus (0.598).<\/li>\n
- Long context<\/strong> (Infinity Bench plus internal TR benchmarks): Thomson 0.753 \u2013 best in class, by a hair over Gemini (0.750).<\/li>\n<\/ul>\n
In their blog post, Hron and Schwarz write, \u201cThomson matches the best. And beats the rest.\u201d Taking the benchmark scores at face value, that is a fair characterization, particularly given that Thomson is a fraction of the size and operating cost of the frontier models it was up against. For a domain-specific model built by a legal information company, this is a notable achievement.<\/p>\n
But the results are more mixed than the headline suggests. Thomson did not win a majority of the rows. Gemini and GPT-5.5 both beat Thomson in the best-known legal benchmark, LegalBench. Anthropic\u2019s Opus edged Thomson on the Harvey agentic benchmark and beat everyone on coding.<\/p>\n
In addition, the \u201cfine print\u201d in the above graphic reveals that the test conditions were not entirely equal. Thomson used test-time scaling \u2014 which gives it more time and computing power. Gemini and Opus ran in the more-powerful reasoning mode, but GPT-5.5 was tested in non-reasoning mode.<\/p>\n
While non-reasoning mode is faster and more efficient for routine tasks, it is typically less capable when handling complex, advanced or multi-step challenges. That suggests that the benchmark may have understated OpenAI\u2019s performance.<\/p>\n
Some Benefit to Thomson<\/strong><\/h3>\nThrough a Thomson Reuters spokesperson, I asked Hron about this. He said that Thomson did, in fact, benefit from test-time scaling on certain tasks, but that the overall effect was minor. \u201cIts grand average increased from 0.787 without test-time scaling to 0.789 with it,\u201d he said, \u201cand the conclusions from the benchmarks would not have materially changed without it.\u201d<\/p>\n
With regard to the use of non-reasoning mode for GPT-5.5, Hron said that the reasoning-mode evaluation took substantially longer to complete, and that those results were not ready in time for publication.<\/p>\n
\u201cWe expect reasoning mode to improve its performance but based on the results we have seen so far, not dramatically,\u201d Hron said. \u201cWe are continuing to run those evaluations and plan to include the results in the technical report.\u201d<\/p>\n
Hron said it is also important to note that reasoning mode and test-time scaling are not necessarily direct analogs.<\/p>\n
\u201cA model operating in a mode described as \u2018non-reasoning\u2019 may still benefit from test-time scaling or other inference-time techniques implemented behind the provider\u2019s API,\u201d he said. \u201cBecause external providers generally do not disclose the full details of those systems, we cannot know precisely which techniques were applied.<\/p>\n
\u201cWe made a best effort to ensure a fair comparison based on the information available and disclosed Thomson\u2019s use of test-time scaling in the interest of transparency.\u201d<\/p>\n
The Retrieval Test<\/strong><\/h3>\nIn a second evaluation, Hron and Schwarz compared Thomson\u2019s performance when integrated with Thomson Reuters\u2019 proprietary content with leading frontier models given unrestricted access to the web.<\/p>\n
The benchmark that drove the headline pits Thomson (technically called Thomson-1-Large) against Gemini 3.1 Pro, Claude Opus 4.8, and GPT-5.5 across three legal benchmarks and four general-domain composites. Thomson took the top score in three of the seven rows:<\/p>\n Legal domain:<\/strong><\/p>\n General domain:<\/strong><\/p>\n In their blog post, Hron and Schwarz write, \u201cThomson matches the best. And beats the rest.\u201d Taking the benchmark scores at face value, that is a fair characterization, particularly given that Thomson is a fraction of the size and operating cost of the frontier models it was up against. For a domain-specific model built by a legal information company, this is a notable achievement.<\/p>\n But the results are more mixed than the headline suggests. Thomson did not win a majority of the rows. Gemini and GPT-5.5 both beat Thomson in the best-known legal benchmark, LegalBench. Anthropic\u2019s Opus edged Thomson on the Harvey agentic benchmark and beat everyone on coding.<\/p>\n In addition, the \u201cfine print\u201d in the above graphic reveals that the test conditions were not entirely equal. Thomson used test-time scaling \u2014 which gives it more time and computing power. Gemini and Opus ran in the more-powerful reasoning mode, but GPT-5.5 was tested in non-reasoning mode.<\/p>\n While non-reasoning mode is faster and more efficient for routine tasks, it is typically less capable when handling complex, advanced or multi-step challenges. That suggests that the benchmark may have understated OpenAI\u2019s performance.<\/p>\n Through a Thomson Reuters spokesperson, I asked Hron about this. He said that Thomson did, in fact, benefit from test-time scaling on certain tasks, but that the overall effect was minor. \u201cIts grand average increased from 0.787 without test-time scaling to 0.789 with it,\u201d he said, \u201cand the conclusions from the benchmarks would not have materially changed without it.\u201d<\/p>\n With regard to the use of non-reasoning mode for GPT-5.5, Hron said that the reasoning-mode evaluation took substantially longer to complete, and that those results were not ready in time for publication.<\/p>\n \u201cWe expect reasoning mode to improve its performance but based on the results we have seen so far, not dramatically,\u201d Hron said. \u201cWe are continuing to run those evaluations and plan to include the results in the technical report.\u201d<\/p>\n Hron said it is also important to note that reasoning mode and test-time scaling are not necessarily direct analogs.<\/p>\n \u201cA model operating in a mode described as \u2018non-reasoning\u2019 may still benefit from test-time scaling or other inference-time techniques implemented behind the provider\u2019s API,\u201d he said. \u201cBecause external providers generally do not disclose the full details of those systems, we cannot know precisely which techniques were applied.<\/p>\n \u201cWe made a best effort to ensure a fair comparison based on the information available and disclosed Thomson\u2019s use of test-time scaling in the interest of transparency.\u201d<\/p>\n In a second evaluation, Hron and Schwarz compared Thomson\u2019s performance when integrated with Thomson Reuters\u2019 proprietary content with leading frontier models given unrestricted access to the web.<\/p>\n<\/a><\/p>\n\n
\n
Some Benefit to Thomson<\/strong><\/h3>\n
The Retrieval Test<\/strong><\/h3>\n
![\"\"](\"https:\/\/i0.wp.com\/www.lawnext.com\/wp-content\/uploads\/2026\/08\/TR8347338_02B_1920x1080-scaled-1-1024x576.jpg?resize=1024%2C576&ssl=1\" "\\"\\"")
<\/a><\/p>\n