Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableRB.RU→ original

Russian platforms charge 9 times world prices for DeepSeek and don't guarantee privacy

Russian platforms sell access to DeepSeek, Claude and GPT at 1.6–9 times the price of foreign aggregators. For a typical workload of 10 million input tokens and 2 million output tokens, DeepSeek V4.1 Flash costs 4,000 ₽ at Yandex AI Studio versus 443 ₽ elsewhere. Neither Yandex AI Studio, GigaChat API nor Cloud.ru commits in its contracts to ZDR mode, meaning no retention of requests and responses after processing. For code and customer data, compare both the price and the written guarantees.

AI-processed from RB.RU; edited by Hamidun News
Russian platforms charge 9 times world prices for DeepSeek and don't guarantee privacy
Source: RB.RU. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceRB.RU ↗
Editor-in-Chief
Claude Opus 5.514.2 thousand ₽ in Russia versus 6.6 thousand ₽ abroad
Claude Sonnet 5.5 and GPT-6.1 Sol7.1 thousand ₽ each versus 3.3 thousand ₽
Key metric #3OpenAI and Anthropic offer ZDR by agreement with the client

Suppose your team has connected DeepSeek, Claude or GPT through a Russian cloud platform, because direct access to foreign providers isn't available. It's convenient, but the bill for that access can run 1.6–9 times higher than what the same models cost abroad. On top of that, the provider may have no written commitment not to store your requests. Nodul, a platform for AI agents, examined the pricing and data-processing terms of 15 model providers. There are two problems: price and privacy.

How much the middleman costs

Because of restrictions on direct access, Russian companies go through domestic platforms that offer work with foreign models. Nodul's analysts used an identical workload: 10 million input tokens (the text you send to the model) and 2 million output tokens (what it replies). A token is a fragment of text, roughly a syllable or a short word, and it is what gets billed.

The gap is largest for DeepSeek V4.1 Flash. At Yandex AI Studio that volume costs 4 thousand rubles; at foreign aggregators it costs 443 rubles. That is nine times more. The gap is smaller for pricier models, but it is still noticeable:

  • Claude Opus 5.5: 14.2 thousand rubles in Russia versus 6.6 thousand abroad, more than double;
  • Claude Sonnet 5.5 and GPT-6.1 Sol: 7.1 thousand rubles each versus 3.3 thousand rubles at international rates.

Across the full set of providers, the markup ranged from 1.6 to 9 times. In the study's examples, the markup hits hardest the cheapest model, which people usually choose precisely because it is inexpensive.

What ZDR is and why it's missing

The second part of the study concerns data. The analysts checked three Russian services: Yandex AI Studio, GigaChat API and Cloud.ru. None of them commits in its public contracts to a full Zero Data Retention, or ZDR, mode. This is the mode in which a provider undertakes to store neither your requests nor the model's responses after processing.

Abroad, the terms vary. OpenAI and Anthropic offer ZDR by agreement, meaning for a specific client. Some aggregators enable it through settings, and elsewhere it works by default. Russian services also use information-protection mechanisms, but according to Nodul, these are not equivalent to full ZDR.

For a request like "rewrite this email," the risk is small. The situation is different for AI agents that work with corporate code, contracts and customer databases: all of that leaves the company with the request. In a related item, RB.RU cites a recent case in which the top manager of a Moscow company was dismissed after using DeepSeek because of the disclosure of a trade secret. The court deemed the employee's actions a gross violation of her job duties.

What to do in practice

Nodul's CEO Alexander Danilov recommends two steps. First, don't run everything through the most powerful model. Route simple tasks, such as sorting support tickets, extracting fields from text or writing short replies, to cheaper models, and reserve the expensive ones for complex reasoning and code. Second, for corporate data, choose services that provide the necessary confidentiality guarantees.

We would add a few checks before choosing a provider:

  • calculate cost on your real token volume, not from the per-million price in the price list;
  • ask the provider for its log-retention terms and how long logs are kept;
  • strip secrets from requests: keys, passwords, client identifiers;
  • decide in advance which data must never be sent to external models at all.

Limitations of the study

It matters who did the counting. Nodul sells its own platform for AI agents, so the company has an interest in showing that alternatives cost more. The calculation covers a single scenario: 10 million input and 2 million output tokens. Volume discounts and individual corporate terms may not have been captured in this comparison. On ZDR, only three Russian services were examined, and only their public contracts. A provider may promise more to a large client. But if guarantees are not on paper, they cannot be relied on in an architecture.

Story

Key Questions Answered

How much the middleman costs?

Because of restrictions on direct access, Russian companies go through domestic platforms that offer work with foreign models. Nodul's analysts used an identical workload: 10 million input tokens (the text you send to the model) and 2 million output tokens (what it replies). A token is a fragment of text, roughly a syllable or a short word, and it is what gets billed. The gap is largest for DeepSeek V4.1 Flash. At Yandex AI Studio that volume costs 4 thousand rubles; at foreign aggregators it costs 443 rubles.

What ZDR is and why it's missing?

The second part of the study concerns data. The analysts checked three Russian services: Yandex AI Studio, GigaChat API and Cloud.ru. None of them commits in its public contracts to a full Zero Data Retention, or ZDR, mode. This is the mode in which a provider undertakes to store neither your requests nor the model's responses after processing. Abroad, the terms vary. OpenAI and Anthropic offer ZDR by agreement, meaning for a specific client.

What to do in practice?

Nodul's CEO Alexander Danilov recommends two steps. First, don't run everything through the most powerful model. Route simple tasks, such as sorting support tickets, extracting fields from text or writing short replies, to cheaper models, and reserve the expensive ones for complex reasoning and code. Second, for corporate data, choose services that provide the necessary confidentiality guarantees.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…