How we scored: 7 criteria and a scale
Each criterion gives 0 to 2 points, with a maximum of 14. We give a point only for what the vendor itself wrote in the product description, terms, prices, or documentation: if it is not on the page, it is 0.
1. Data: 2 if documents do not leave the firm’s server or a cloud in Russia; 1 if data is abroad but the vendor promises in writing not to train models on it; 0 in all other cases.
2. Access: 2 for an open license with no country limits; 1 if the terms do not exclude Russia but do not promise access; 0 if Russia is not on the vendor’s country list or there is no list.
3. Citations: 2 for a link to a legal database with quote checking, 1 for a link to an uploaded document or a website.
4. Russian law: 2 for a stated database of Russian laws and case law, 1 for partial coverage.
5. Volume: 2 for a context from 1 million tokens or thousands of documents, 1 for several documents without a stated volume.
6. Price: 2 for a published price or a free license, 1 for a per-user plan after a request, 0 for a private offer.
7. Control: 2 if the system, log, and access are fully with the firm; 1 for admin settings, an audit log, single sign-on, or a SOC 2 audit at the vendor.
For a firm in Russia, the law decides the first two criteria. Before a cross-border transfer of personal data, the operator notifies Roskomnadzor, the Russian data regulator (Article 12(3) of 152-FZ, Russia’s personal data law); citizens’ data is stored in databases in Russia (Article 18(5)); and client information is covered by attorney-client privilege under Article 8 of the Law on Advocacy.
A firm in the US would rank the same tools differently. Our ranking answers the question of a firm in Russia.
AI ranking for lawyers: points out of 14
Ranked by total points. In brackets: data, access, citations, law, volume, price, control.
1. Qwen3 on your own server: 10 (2, 2, 0, 0, 2, 2, 2).
2. DeepSeek V4 on your own server: 10 (2, 2, 0, 0, 2, 2, 2).
3. Llama 4 on your own server: 9 (2, 1, 0, 0, 2, 2, 2).
4. Claude: 7 (1, 0, 1, 0, 2, 2, 1).
5. CoCounsel Legal: 7 (1, 0, 2, 0, 2, 1, 1).
6. Gemini: 6 (1, 0, 1, 0, 2, 2, 0).
7. ChatGPT: 5 (0, 0, 1, 0, 2, 2, 0).
8. DeepSeek, cloud service: 5 (0, 1, 0, 0, 2, 2, 0).
9. Harvey: 5 (1, 0, 1, 0, 2, 0, 1).
10. Spellbook: 4 (1, 0, 0, 0, 1, 1, 1).
11. Lexis+ AI: 3 (0, 0, 2, 0, 1, 0, 0).
12. Luminance: 2 (0, 0, 0, 0, 1, 0, 1).
Everyone scores 0 on Russian law: none of the twelve options states a database of Russian laws and case law. It is connected separately, inside the firm’s own systems.
Legal AI services: Harvey, CoCounsel, and others
Legal services are strong in citations and volume, but none names Russia among its regions.
1. CoCounsel Legal, 7 points. According to Thomson Reuters, answers rely on Westlaw, Practical Law, and the firm’s documents, citations open for checking, table analysis handles up to 10,000 documents on 100 questions, and client data does not go into training. Plans for 1, 2, and 3 years are sold through the US website.
2. Harvey, 5 points. Data is processed in the EU or Switzerland, the US, or Australia; the contract forbids model vendors to train on client data; and there are audit logs and single sign-on. Answers rely on the client’s sources, storage is built for thousands of documents, and there is no price on the website. According to the company, 80% of firms use Harvey among the Am Law 100.
3. Spellbook, 4 points. A Word add-in for contracts: zero data retention rules out training on the data, there is a SOC 2 Type II audit, a 7-day free trial, and a per-user price after a demo. The data storage location and sales countries are not stated.
4. Lexis+ AI, 3 points. LexisNexis launched it in 2023 for clients in the US: citations lead into the LexisNexis database and are checked with Shepard’s. In the Stanford test, Lexis+ AI still made errors in more than 17% of answers. We did not find the price or data storage terms publicly.
5. Luminance, 2 points. A contract service used by more than 1,000 companies, with a SOC 2 Type II audit. On AWS Marketplace it is sold only through a private 36-month offer with payment per document; the data storage location is not stated.
General models lawyers use: ChatGPT, Claude, and others
General models offer a large context and open prices per 1 million tokens. Three vendors out of four do not include Russia in their country lists, and the fourth stores data in China.
1. Claude, 7 points. A context of 1 million tokens, about 555,000 words; the Citations feature returns the exact document passage under each claim; data on commercial plans does not go into training by default, and sending feedback is managed in the organization’s settings. Russia is not on Anthropic’s list of regions.
2. Gemini, 6 points. In the paid API, requests are not used to improve products. In the free tier, human reviewers may read them, and Google itself asks you not to send confidential or personal data there. Google Search gives links to websites, not to a legal database; Russia is not on the Gemini API list of regions.
3. ChatGPT, 5 points. The context of the flagship API model is 1.05 million tokens, file search links to the uploaded file, and prices are published. Russia is not on OpenAI’s country list, and in The New York Times lawsuit, a court in 2025 ordered OpenAI to keep answer logs, including deleted chats.
4. DeepSeek as a cloud service, 5 points. Under its privacy policy, the company stores personal data in China and uses it to train models; the terms follow Chinese law and do not promise access in specific countries. For client documents, this is a cross-border transfer and disclosure of privileged information to a third party.
Open models on your own server: Qwen, DeepSeek, Llama
An open model is downloaded with its weights and run on the firm’s server or in a cloud in Russia. There is no cross-border transfer, and the log and access stay with the firm, hence 2 points each for data and control.
1. Qwen3–235B-A22B, 10 points. Apache 2.0 license, a context of 262,144 tokens that extends to 1,010,000, with 22 billion of 235 billion parameters active.
2. DeepSeek V4 Pro, 10 points. The weights of the cloud service are under the MIT license, with a context of 1 million tokens. The model is large: 1.6 trillion parameters, 49 billion of them active, so a server for it costs more.
3. Llama 4, 9 points. The license is free worldwide but requires you to follow trade restriction laws: 1 point for access. Llama 4 Scout has a context of 10 million tokens, but Russian is not on the list of officially supported languages.
The zeros for citations and Russian law are fair for open models: without setup they answer like a general chatbot, and such chatbots made errors on 58–82% of legal questions in an earlier Stanford study. The rollout adds a law database, the firm’s archive, and citation checks.
What a Russian firm can take from the ranking
The conclusion for a firm in Russia: client documents go to a model inside its own systems, and foreign services serve as a model of the features you need.
1. Take the requirements from legal services: a link to every source, quote checks against a database, an audit log, and a ban on training on your data.
2. Install an open model on your own server or in a cloud in Russia, and connect the firm’s archive and the texts of laws.
3. Test the system on closed cases. The test procedure and 12 criteria are on the page about how to choose AI for lawyers.
Since 1 September 2026, Federal Law 243-FZ has been in force: the Russian government decides when sovereign and national models are mandatory. A model inside the firm’s systems can be replaced if the requirements reach lawyers.
You can start with one process, such as contract review: a pilot from $1,499 for 14 days.