Which AI works in Russia in 2026: GigaChat, Alice AI, Cotype and other ChatGPT alternatives compared

Why "just use ChatGPT" doesn't work in Russia

For a Russian business, ChatGPT has three practical problems: access and payment require a VPN and a foreign card, customer data goes to overseas servers (a compliance question under Russia's 152-FZ law), and there are no integrations with Russian systems — 1C, MoySklad, marketplace accounts. So companies are switching to Russian assistants: they work without a VPN, accept payment in rubles, and keep data in Russia.

The demand is no longer experimental. 71% of large Russian companies use generative AI in at least one business function (a study by Yakov & Partners and Yandex, December 2025). Russia's generative AI market for 2025 is estimated at roughly 58 billion rubles, with a forecast of 778 billion by 2030 (Just AI and Onside). A CNews review from February 11, 2026 adds the detail that matters most: 64% of companies use AI specifically as an assistant — under mandatory human control, not as an autonomous executor.

Four categories of "AI for business" — and why they're not the same thing

A year ago there were three. In 2026 a fourth appeared between the general-purpose assistant and the finished agent: the builder. The difference decides how much work actually comes off your plate — and how much lands on it.

1. General-purpose assistant. Answers, writes copy, summarizes documents, runs RAG over a knowledge base. GigaChat, Alice AI, Cotype. Applying the result inside your systems is still on you.
2. Open model in your own perimeter. A model you deploy yourself: T-Pro 2.0 and T-Lite from T-Technologies under Apache 2.0, GigaChat 3 and 3.1 weights under MIT, Cotype Light 3 (9B parameters, about 18 GB of VRAM — a single accelerator). Maximum control over your data, but you need a developer and hardware.
3. Agent builder. An environment where you assemble the agent: GigaCowork and GigaChat Business from Sber, MWS AI Agents Platform from MTS, Agent Atelier with MCP Hub from Yandex, Evolution AI Agents from Cloud.ru, the joint Selectel and Just AI platform. It is sold as "AI agents," but what you buy is a builder: the logic, the connectors, the result checks, and the responsibility for them are yours.
4. Vertical agent. Built for one specific niche and performs the action inside your system itself: calculates economics, changes a price, replies to a review, builds a shipment. You set the rule in plain words — it works to that rule.

Comparison: Russian AI services for business

As of August 18, 2026. Data is taken from vendors' own pages and public announcements (sources at the end). Where a vendor publishes "on request" only, the table says so.

What this shows. General-purpose assistants got better and cheaper over the year, but applying the decision inside your accounts is still your job. Open models trade development effort for control over data — and here 2026 brought the real change: GigaChat's weights are published under MIT, so "a Russian assistant" and "an open model in your own perimeter" stopped being opposites. Agent builders are the loudest category of the year, but buyers often miss that they are buying a construction site rather than a result. And a vertical agent takes over the routine of one niche completely — but only its own niche.

An agent builder is not an agent

This is the fork of 2026, and it is where budgets get lost.

When Sber, Yandex, MTS, and Cloud.ru say "AI agents," what they technically offer is a platform: a set of models, function calling, connectors, role-based access, a scenario editor. All of that is honest and useful — but there is no agent in it that starts working on Monday. There is an environment in which your analyst or contractor will build one: describe the steps, connect the systems, invent the checks, handle the cases where a system answered something unexpected. That is a project, not a subscription.

The difference shows up in a single question to the vendor: "who is accountable for the number in the answer being correct?" A builder has no answer — you write the method. A vertical agent has it built in, and you can test it.

A separate case is assistants embedded in business systems. In Bitrix24's Cosmos release (April 2026), CoPilot became BitrixGPT 5, and an agent named Marta appeared alongside it, connecting to external systems over MCP. In amoCRM, the April 18, 2026 conference introduced a built-in agent named Amma that reviews deals and prompts the sales rep. This is convenient and nearly free if you already live inside that CRM — and completely bounded by its perimeter.

An assistant writes. An agent does.

Ask a general-purpose assistant to "raise the price on an unprofitable item" and it will explain how to calculate that and draft the text — while you open the account, work out the margin on every SKU, and change the bid. A vertical agent does it itself: pulls cost from your accounting system (1C or MoySklad), finds the SKUs leaking margin, recalculates the price backwards from target profit, replies to the review in the brand's voice, and builds the shipment.

In 2026 this border gained new neighbors, and they deserve an honest mention. Wildberries launched an AI "Assistant" for sellers, available on the Jam subscription, showing sales, stock, demand trends, and search queries. Ozon is testing a "Smart assistant" in the seller account with a limited group of sellers. Yandex Market gave sellers an assistant chat back in mid-2025. All three are assistants in the precise sense: they show numbers for their own platform. They do not change prices, do not answer a buyer's review, do not reconcile account revenue against shipments in your accounting system, and cannot see the marketplace next door.

At Samreshuuu, "vertical" doesn't mean one model that knows about marketplaces — it means a set of named skills, each with its own method: Unit economics by SKU breaks platform reports down to net profit per item, Marketplace pricing works the price backwards from a target margin, Purchase forecasting maintains safety stock and the reorder point, and Marketplace reviews and questions answers buyers through official APIs. That method is exactly what a general-purpose assistant lacks — not because the model is weaker, but because it has nowhere to get your numbers or the industry formula. Cost, meanwhile, comes from your accounting system — 1C or MoySklad — not from the product card.

One prompt shows the difference:

Data and the law: what changed in summer 2026

On July 26, 2026, Russia's first law on artificial intelligence was signed — on supporting the development of AI technologies (bill No. 1271570-8). Its main provisions take effect on September 1, 2026; the rules on content labeling and government authority follow on March 1, 2027, with a transition period running to 2032.

What the law does: it regulates large foundation models from 1 billion parameters up, creates two privileged categories — sovereign (fully Russian development, Russian data centers) and national (foreign components allowed under an open license) — grants their developers access to state support, and permits training on publicly available copyrighted works without separate consent from the rights holder. Disclosure of AI-generated origin is voluntary.

What the law does not do — and this matters more when choosing a service: it does not regulate personal data. That still belongs to 152-FZ, where 2026 raised fines by roughly an order of magnitude: a customer-base leak now costs millions of rubles, and a repeat biometric leak carries a turnover-based fine. Terms of service deserve a separate read: Fontanka's analysis of the new law points out that GigaChat's user agreement assumes a perpetual and irrevocable license to the content you upload.

The practical takeaway: "it's a Russian service, so the data is fine" is a simplification. The right question is different — where does the data physically sit, what do the terms say about rights to what you upload, and can the processing move into your own perimeter? In 2026 the last question can finally be answered "yes" without giving up quality: the open weights of GigaChat 3 and Cotype Light 3 cover exactly that scenario.

If you sell to the public sector or bid in procurement under the national regime, add the Russian software registry to the list: it now has a "trusted software" track and separate requirements for generative models, with compatibility requirements phasing in from September 1, 2026.

How to choose an AI assistant for business: a checklist

1. Do you need copy, control over data, a platform, or a finished result? Copy — a general-purpose assistant. Your own perimeter and 152-FZ — an open model. Many different processes and a team of your own — an agent builder. Routine off your plate in one niche without a project — a vertical agent.
2. Who is accountable for the number being right? With a builder, you are. With a finished agent, the vendor is — and you can test that before buying: ask to see how the agent names the data slice and admits an incomplete result.
3. Where must the data stay, and what do the terms say? For personal and financial data, check Russian residency, the on-prem option, and the rights to uploaded content. The new AI law does not settle this — it governs models, not personal data.
4. Do you have developers? Open models and builders require them; ready-made assistants and vertical agents don't.
5. Do you need integrations with your systems? 1C, MoySklad, CRM, marketplace accounts — with models and builders you build these yourself; a vertical agent ships with them.
6. One service or a stack? Add up subscriptions, implementation, and the cost of manually reconciling between tools. With a builder, implementation is its own budget line — usually larger than the subscription.

Frequently asked questions

Is there a Russian AI like ChatGPT? Yes. GigaChat from Sber and Alice AI from Yandex are general-purpose assistants of the same kind: chat, copy, document analysis, code, RAG over your files. Cotype 3 from MWS AI is the corporate version that runs on-prem. They work without a VPN, GigaChat and Alice AI bill in rubles, and the data stays in Russia. In global knowledge and non-Russian languages they generally trail ChatGPT and Claude; in Russian-language tasks and local context they are the ones optimized for it.

Which AI works in Russia without a VPN? Every Russian service in the table: GigaChat, Alice AI, Cotype, plus open T-Pro 2.0, the GigaChat 3 weights and Cotype Light 3 if you run them yourself. ChatGPT, Claude and Gemini need a VPN and a foreign card to pay, and there are no integrations with 1C, MoySklad or marketplace accounts. For work inside Ozon, Wildberries, Yandex Market or 1C, a vertical agent such as Samreshuuu connects to those systems directly and needs no VPN either.

What can replace ChatGPT for business in Russia? Russian assistants: GigaChat from Sber and Alice AI from Yandex work without a VPN, accept payment in rubles, and keep data in Russia. For your own perimeter, use open T-Pro 2.0 under Apache 2.0, Cotype Light 3, or GigaChat 3 weights under MIT. For a specific niche such as marketplaces — a vertical agent.

Which Russian AI assistant suits a small business? For drafts, copy, and document analysis, GigaChat or Alice AI on pay-as-you-go is enough — GigaChat's free tier for individuals now runs up to 365 million tokens. If the business is on marketplaces and the goal is to take account routine off your plate, you need an agent that opens Wildberries, Ozon, and Yandex Market itself and calculates profit from cost.

What's the best neural network for business in 2026? There's no single "best": GigaChat 3.5 Ultra and Alice AI are strong general-purpose assistants, T-Pro 2.0 and Cotype Light 3 are for your own perimeter, and for e-commerce a vertical agent does the work in the account. The choice depends on whether you need copy, control over data, or actions.

How does an AI agent differ from an agent builder? A builder is an environment in which you assemble the agent: the logic, connectors, checks, and accountability for the result are yours. A finished vertical agent arrives with the method inside. A simple pre-purchase test: ask how long the first working scenario takes. With a builder the answer is measured in weeks of implementation; with a finished agent, in minutes of connection.

Are these all ChatGPT alternatives? General-purpose assistants (GigaChat, Alice AI, Cotype from MWS AI) — yes, direct alternatives for text and analysis. Agent builders and vertical agents — no: the first are platforms, the second are executors inside a specific system.

Is it safe to give AI access to company data? Yes, if the data stays in Russia or in your own perimeter, the connection goes through an official API (a token, not a login and password), and the agent has an "approval" mode. Check the service's user agreement separately — rights to uploaded content can be broad at Russian services too.

---

Last updated: August 18, 2026.

Sources: GigaChat — developers.sber.ru (individual and corporate tariffs effective February 1, 2026; changelog entries of July 15 and 22, 2026), CNews and Sostav on the GigaChat 3.5 Ultra announcement (July 6, 2026), Habr / SberDevices on GigaChat 3 Ultra Preview, weights at huggingface.co/ai-sage; GigaCowork — Habr and Kommersant (May 19, 2026); Alice AI and Yandex AI Studio — aistudio.yandex.ru (Agent Atelier, MCP Hub), Vedomosti on Alice AI LLM Flash (May 28, 2026); Cotype — Vedomosti and itWeek on Cotype Pro 3 and Light 3 (July 15, 2026), TAdviser on the MTS AI to MWS AI rename; MWS AI Agents Platform and AI Force — CNews (March 4 and May 21, 2026); Cloud.ru — cloud.ru/blog (Evolution AI Factory); Selectel and Just AI — CNews and selectel.ru (June 25, 2026); Bitrix24 and amoCRM — Cosmos release and Amoconf 2026 coverage; marketplace assistants — RB.ru and CNews (Ozon, August 4, 2026); AI law No. 1271570-8 — lidings.com and Fontanka (July 30, 2026); statistics — Yakov & Partners and Yandex (December 2025), CNews (February 11, 2026), Just AI and Onside.