RGBvision RGBvision
Back
aimarket

DeepSeek-R1: what a cheap open model changes in AI automation

DeepSeek released an open reasoning model that costs a fraction of OpenAI o1 via API. Where it helps a business, where it falls short, and what 152-FZ says.

Published: 2025-01-28

On January 20, the Chinese company DeepSeek released R1, a reasoning model that it says matches OpenAI o1 in math, code and logic while costing roughly 27 times less through the API. The weights are open under the MIT license. The short answer for a business: tasks where a model has to think are now much cheaper, but R1 is not a ready-made brain for an AI agent, and sending it customers' personal data from Russia triggers the cross-border transfer rules of 152-FZ, Russia's personal data law.

What shipped on January 20

DeepSeek published several things at once:

  • DeepSeek-R1: 671 billion parameters, 37 billion of which are active per token. Available in DeepSeek's chat, through the API (model deepseek-reasoner) and as open weights on Hugging Face.
  • MIT license: commercial use, modification and distillation are allowed. API outputs can also be used to fine-tune your own models.
  • Six distilled models: R1 was used to train smaller models based on Qwen 2.5 and Llama, at 1.5, 7, 8, 14, 32 and 70 billion parameters. The Llama-based versions also fall under the Llama license.
  • A technical report describing the training, with a candid section on limitations that we come back to below.

Launch API prices, in US dollars per 1 million tokens:

ModelInputOutput
DeepSeek-R1 (deepseek-reasoner)0.55 (0.14 on cache hit)2.19
OpenAI o11560

On DeepSeek's own tests, R1 scores 79.8% on AIME 2024 against 79.2% for o1, and 97.3% on MATH-500 against 96.4%; o1 is slightly ahead on Codeforces. These are the developer's numbers, and independent evaluations are only starting.

The market reacted sharply. On January 27, Nvidia shares fell 17% and the company lost $589 billion in market value, the largest one-day drop in market history. The same day, the DeepSeek app topped the free charts in Apple's US App Store, and the company temporarily limited new sign-ups, citing large-scale malicious attacks on its services.

What changes for automation

Reasoning tasks got cheaper. Models like o1 do well where the job is not just to answer but to weigh conditions against each other: parse a specification with complex discounts, check a supplier's price list against the contract, classify a request with ambiguous wording. Before R1, many of these scenarios hit a cost ceiling.

A rough example: a request where the model reads 3,000 tokens and writes 1,500 (the reasoning is billed as output tokens too). For 1,000 such requests at the prices above, o1 costs about $135 and R1 about $5. This is a calculation from the price lists, not a measurement: actual token usage depends on the task and the model.

You can now host the model yourself. Open weights under MIT can run on your own servers, so the data never leaves for a third-party provider.

For first-line support it changes little. A reasoning model thinks for seconds, sometimes minutes, before it answers. A customer in a messenger needs a fast, accurate answer from the knowledge base, not a chain of reasoning. We solved this in our own product, DialIQ: an AI operator answers customers in Telegram, WhatsApp, SMS and website chat according to the company's role and knowledge base, and hands complex conversations to a person. In this scenario, speed and predictability matter more than scores on olympiad problems.

The catches

DeepSeek's own technical report lists R1's limitations, and for a business they matter more than benchmark records.

  • Weaker agent skills than the base model. The authors say R1 falls short of their own DeepSeek-V3 in function calling, multi-turn dialogue and JSON output. That is exactly what an agent reading and writing to a CRM relies on. A sensible setup: R1 handles the hard step, and another model performs the actions in your systems.
  • Russian is not a priority. The model is optimized for Chinese and English and may mix languages when answering in others. Quality on your Russian documents has to be tested on your own examples.
  • Your usual prompts may not fit. According to the report, few-shot examples degrade R1's results, and the authors recommend stating the task without them. Prompts tuned for other models will need rework.
  • Hosting the full model is expensive. By our estimate, 671 billion parameters need hundreds of gigabytes of GPU memory, meaning a server with several data-center accelerators. The 7–14B distilled models run on far more modest hardware but are weaker: high math scores do not guarantee quality on your contracts.
  • A young service under peak load. The sign-up restriction shows the API can become unavailable at the worst moment. A process that depends on a single provider stops when it does.

Personal data and 152-FZ

DeepSeek is a Chinese company. If a request to its API contains personal data, such as a customer's name, phone number or a message containing that data, this is a transfer of personal data to a foreign legal entity on the territory of a foreign state, which 152-FZ defines as cross-border transfer (Art. 3(11)). Since March 1, 2023, such transfers have been governed by Article 12 of 152-FZ as amended by Federal Law 266-FZ:

  • before starting the transfer, the operator must notify Roskomnadzor, Russia's data protection regulator, separately from the notification of intent to process personal data (Art. 12(3));
  • before filing that notification, the operator must obtain from the foreign recipient information on its data protection measures and the conditions for ending processing, and, if the recipient's country is neither a party to the Council of Europe Convention for the Protection of Individuals with regard to Automatic Processing of Personal Data nor on the regulator's list of countries with adequate protection, information on that country's legal framework (Art. 12(5));
  • to such countries, once the notification is filed, data may not be transferred until the review period has expired, except where the transfer is needed to protect the life, health or other vital interests of the data subject or other persons (Art. 12(11)).

These rules apply to any foreign API, not just DeepSeek. The shortest path is not to send personal data abroad at all: pilot on tasks without it, and for tasks with it, use a model on your own server in Russia or a Russian service such as YandexGPT or GigaChat. If a foreign model is unavoidable, prepare the notification procedure together with a lawyer.

What this means for your business

Cost is no longer the main barrier for tasks where AI has to reason, and in our forecast not only at DeepSeek: other providers will have to respond to these prices. The winners are companies where the model is a replaceable part, not the foundation. Today R1 is cheaper; tomorrow something cheaper or better at Russian will appear.

For tasks involving personal data, R1 through the API is not the default choice, and for agents working in CRM and accounting systems it does not replace the main model. Its place is in individual hard steps of a process.

What to do

  1. Pick one or two tasks without personal data that require logic: checking estimates and specifications, reconciling price lists, parsing technical texts.
  2. Collect 50–100 real examples with correct answers and run R1, your current model and one or two distilled models on them. Look at accuracy, response language and latency.
  3. Calculate the cost on your real volume, including reasoning tokens, and compare it with what this work costs now.
  4. Separate the model from the agent's logic. Call the model through a dedicated layer where it can be swapped by configuration rather than by rewriting the agent. Plan a fallback model in case the provider goes down.
  5. Decide in advance how to handle personal data: your own server, a Russian provider, or the cross-border transfer procedure.

A pilot like this on one process takes 2–4 weeks. If you need help choosing a model and designing the architecture, see how we build AI agents and automation.

Sources