DeepSeek

There are two things to consider: whether you need a chat without payment or workarounds, and what you plan to upload to it.

Who it is for. For people who want strong AI for free and without access hassles: the chat opens directly from Russia, and you can ask the first question without registration. According to the a16z report, in March 2026 DeepSeek held 29 % of Russia's chatbot market, more than ChatGPT.

Who else. Developers and companies that cannot send data abroad: V4 weights are open under MIT, so Yandex and Cloud.ru sell the same model in rubles, and you can run it on your own hardware if needed.

Who it is not for. Those who work with other people's data and contracts: correspondence under the vendor's policy is stored in China (10.02.2026). And those who need flawless Russian: in reviews from September 2026, the model switches to German or Kazakh in long answers.

A fit: section 3 gets you to your first question in three steps. Choosing between it and ChatGPT: the fork is in section 7.

What vibecoding isStarting from scratch: how to set tasks for agents with no coding experience

Three chat modes, search, and files in one window, with two proprietary V4 models under the hood.

Instant and Expert. In Instant mode, the model answers right away; in Expert mode, it reasons first: in the official store screenshot, it “thought for 52 seconds” about a sum-of-squares task. Since 13.08.2026, Expert mode has used V4 Pro, and you can choose from three reasoning-depth levels.

Search and images. The Search button pulls in websites and shows how many pages it found and read. Vision mode analyzes photos and screenshots, and documents can be dropped into the chat: in the store screenshot, it is a 1.2 MB text file containing a novel, whose plot the model summarizes.

Under the hood. Two proprietary V4-family models with a one-million-token context window, enough for several thick books in one conversation. V4 Pro has 1.6 trillion parameters, V4 Flash 284 billion; both have weights open under the MIT license, so other clouds run them too.

For developers. The API is compatible with OpenAI and Anthropic formats, so DeepSeek can replace their models in Claude Code, GitHub Copilot, or OpenCode by changing the endpoint. On the same 13.08.2026, the vendor launched DeepSeek Harness, its coding agent in developer-preview status, with a separate catalog page.

What was actually done. On Habr, the M.Video team gave five models a task to find the longest palindrome in a string (26.11.2025): DeepSeek got 10 out of 10, but its code nearly matched answers from Grok and Alice, so the author suspected a memorized template. We have not run DeepSeek on the task ourselves; the numbers come from official screenshots and third-party tests.

You can use it from Russia, and the regulator officially confirms it.

Works from Russia. Open chat.deepseek.com: on 24.05.2026 and 22.07.2026, Roskomnadzor said it had imposed no restrictions, while guides from summer 2026 agree that the chat opens without setup and answers in Russian. The vendor's homepage and documentation are served from Russian nodes in Moscow and St. Petersburg (checked 05.09.2026); the chat itself is hidden behind a bot filter and cannot be tested this way. A step-by-step guide to access, apps, and issues is collected in DeepSeek in Russian and from Russia.

Step 1. Ask your first question without registering; the basic chat works anonymously.

Step 2. Create an account for history and files: sign in with email and password, Google, or Apple. The registration screen from a foreign address says that registration in this region is available by email only (checked 05.09.2026).

Step 3. Choose a mode for the task: Instant for a quick answer, Expert for a complex one, and turn on Search if you need fresh data from the web.

Apps. iPhone: the app is available in the Russian App Store, version 2.4.4 from 03.09.2026, with a Russian interface and a 4.1 rating from 8.6 thousand ratings. Android: the official package is in Google Play, with 50 million downloads worldwide. RuStore has build 2.2.2 from July 2026, and the store listing does not make clear whether the vendor uploaded it. DeepSeek has no desktop apps, only the browser.

Stability. There is no ban, but there are outages. On 24.05.2026, the service would not open for users in Russia; outage trackers counted 21–22 thousand reports in a day. September reviews repeatedly mention “server busy” and “frequent messages” errors.

Steps checked 05.09.2026 using live sign-in windows at chat.deepseek.com, App Store and Google Play listings, a test from Russian nodes, and Roskomnadzor statements in Interfax and Vedomosti

The model is below the top, the audience is at the top, and in Russia it is first among foreign options.

Artificial Analysis index. V4 Pro scores 53 points and ranks 39th overall (05.09.2026); the leader Claude Fable 5.1 scores 65, while the open Kimi K3 and GLM-5.3 score 60 each. But an index task costs $0.27 with DeepSeek versus $0.84 with Kimi K3, cheaper than every comparable open model. Not first on points, but cheaper than everyone nearby.

Vendor claims. In the V4 Pro release notes (13.08.2026): Terminal Bench 2.1 at 87.9, DeepSWE at 62.7. Some benchmarks are internal, with no independent reproduction, so treat the figures as self-reported.

Audience. Similarweb counted 315 million visits to chat.deepseek.com in July 2026, placing it fifth among AI chats worldwide. Russia provides 9.7 % of traffic, second only to China.

Demand in Russia. “Dipsik” gets 1.6 million searches a month, “deepseek” another 145 thousand (Wordstat, 05.09.2026). According to an a16z report (March 2026), DeepSeek has 29 % of Russia's chatbot market versus 27 % for ChatGPT. Its place among assistants in our catalog is shown below the card title, and the full tool ranking is on the main page.

The chat is entirely free; only developers pay, and since 16.08.2026 they pay noticeably more.

Free entry. The chat has no paid plan: the website lists API prices only, and the App Store has no in-app purchases (05.09.2026). The vendor does not publish limits; during heavy load, the chat returns a “server busy” error.

API by hour. The price per million tokens depends on the time of day. V4 Flash charges $0.22 for input and $0.66 for output off-peak, and twice as much at peak; V4 Pro charges $0.66 and $1.98 off-peak. On weekdays, peak hours run from 01:00 to 04:00 and from 06:00 to 10:00 UTC, which is morning and early afternoon in Moscow. Reusing previously sent context costs $0.007, thirty times less.

What changed. Before 16.08.2026, V4 Flash output cost $0.28; now it is $0.66 off-peak and $1.32 at peak, 2.4 and 4.7 times more expensive (Habr, 14.08.2026). Cached V4 Pro input rose by up to 12 times, and Habr concluded: the era of the nearly free DeepSeek API is over.

Paying from Russia. Russian cards and Mir do not work in the API dashboard; you need a foreign card or an intermediary, with options collected in how to pay for AI services from Russia. The ruble route without intermediaries: the same models are sold by Yandex AI Studio and Cloud.ru, 9–17 times more expensive than the direct API, but with a contract for a Russian legal entity and data stored in Russia.

Three questions: where correspondence is stored, what the model stays silent about, and when it makes mistakes.

Data in China. Under its privacy policy dated 10.02.2026, the vendor collects, processes, and stores data in China, with a company in Hangzhou responsible for it. Conversations are used to train models; opting out is stated as a user right, but the policy does not say where to enable it. The rule is simple: do not upload contracts, personal data, or work files to the chat. Need the same model without sending data abroad? Choose V4 from Yandex and Cloud.ru, or run the weights on your own hardware.

Censorship. In ellamind's independent test across 168 topics sensitive to China, DeepSeek V3.2 answered honestly in only 19 % of cases, while Kimi K2.5 did so in 99 % (February 2026). There is no such measurement for V4.

Russian breaks down. Four App Store and RuStore reviews from 01-04.09.2026 complain about the same thing: midway through a long response, the model switches to German, Arabic, or Kazakh. The same defect is logged in DeepSeek's GitHub bug tracker. On the Russian-language MERA benchmark, V3 scored 0,677 versus GigaChat 3 Ultra (16.02.2026); there is no measurement for V4.

Incidents. In January 2025, Wiz researchers found an exposed DeepSeek database containing chat histories and API keys; the vendor closed it after notification. On 21.07.2026, search engines indexed conversations users had themselves shared via public links; private chats were unaffected. On 23.02.2026, Anthropic accused DeepSeek, Moonshot, and MiniMax of training on Claude responses through thousands of fake accounts; the companies did not respond to CNBC's request.

Government restrictions. Under the NDAA law (signed 18.12.2025), the US Department of Defense must remove DeepSeek models from its systems and prohibit its contractors from using them. Italy restricted processing of its users' data from 30.01.2025, and government agencies in Australia and South Korea banned the service on work devices in February 2025. There are no restrictions in Russia; Roskomnadzor confirmed this twice in 2026.

The choice is not "who is smarter," but "how much access costs and what you pay for it."

Access. DeepSeek is available directly from Russia, without registration or a card. ChatGPT is not directly available from Russia; the Plus subscription costs $20 per month and cannot be paid for with a Russian card.

Smarts. On the Artificial Analysis index, GPT-6 Astra scores 61 versus 53 for V4 Pro (05.09.2026). According to AA's measurement, it is slower than average and verbose, but an index task costs it $0,27, and the same model is available in open weights.

Data. DeepSeek conversations are stored in China, while ChatGPT's are stored in the US. For work documents from Russia, both services are foreign; the only difference is jurisdiction.

Who else is nearby. Kimi stands out with long documents and a swarm of agents, while GigaChat beats DeepSeek in Russian on the MERA benchmark and stores data in Russia.

In three years, the hedge-fund lab went from the open V3 model to a chat used more often than ChatGPT in Russia.