In This Article
The year 2025 began with a wave of interest in the Chinese AI model DeepSeek. Its emergence triggered a collapse in the technology stock market and nearly caused heart attacks among top managers of competing projects.
What is DeepSeek?
DeepSeek is a Chinese company specializing in the development of artificial intelligence (AI) and big data technologies. The primary focus of DeepSeek is on creating solutions for data analysis, machine learning, and process automation. The company develops tools and platforms that help businesses and organizations effectively utilize data for decision-making, optimizing operations, and improving products and services.
DeepSeek is also actively involved in natural language processing (NLP), computer vision, and other areas of AI to create innovative products such as intelligent assistants, recommendation systems, and analytical platforms. Their technologies are applied across various industries, including finance, healthcare, retail, and logistics.
The company has a chatbot named DeepSeek. This program is an AI that operates based on NLP. It is available both in browser form and as a mobile application. The language model understands any languages.
Here’s what the chatbot says about itself:
"I am an artificial intelligence created by DeepSeek. My main task is to assist you with questions, provide information, solve tasks, and engage in dialogue on various topics. I work based on natural language processing (NLP) and machine learning technologies, allowing me to understand your requests and generate responses."
The company was founded in May 2023, and the first version of the DeepSeek chatbot was released that same year. However, it wasn't until January 2025 that DeepSeek gained widespread attention.
Introduction
Architecture
DeepSeek-V3 adopts a Multi-head Latent Attention (MLA) architecture, which enables efficient inference and achieves performance comparable to leading closed-source models. The model also employs a DeepSeekMoE architecture, which facilitates economical training and encourages load balancing. The MLA architecture is designed to achieve efficient inference by reducing the computational complexity of the model, while the DeepSeekMoE architecture enables cost-effective training by minimizing the number of parameters.
Training Framework
The training framework of DeepSeek-V3 is designed to achieve efficient inference and cost-effective training. The framework adopts a full computation-communication overlap strategy, which enables the model to achieve efficient inference and reduces training costs. The framework also employs a cross-node MoE training strategy, which enables the model to achieve efficient inference and cost-effective training on extremely large scale models.
How to Use DeepSeek v3
To start using DeepSeek, you need to create an account. Registration requires any email address and a password. The program will send a verification code to the specified address.
From this point onward, users can begin working with the AI model without any payment required.
Main Competitor of ChatGPT in Large Language Models
ChatGPT from American OpenAI revolutionized the chatbot market at its time. No other program offered such capabilities before it.
"ChatGPT first appeared on November 30, 2022, when OpenAI released it as a public beta based on the GPT-3.5 model. Later, in March 2023, ChatGPT Plus with GPT-4 was introduced. Since then, it continues to evolve, enhancing its capabilities," is how the chatbot describes itself.
As of January 2025, several models are available in both the app and web versions of ChatGPT. The most advanced is o1 (omni). According to the developers, it significantly surpasses its predecessors.
On January 20, 2025, a new model called DeepSeek-R1 was released. It is built on an outdated model, deepseek v3. It also includes elements from deepseek coder v2 (this version of the neural network was designed for coding tasks).
The emergence of DeepSeek-R1 caused panic in the stock market and led to a decline in AI sector stock prices. Additionally, the release negatively impacted cryptocurrency values. From that day on, a battle began between the two giants: Chinese DeepSeek and American OpenAI.
The fact is that DeepSeek turned out to be cheaper than ChatGPT. The company spends less money on developing and maintaining the chatbot. Moreover, the subscription cost for the Chinese neural network is lower. The rates for ChatGPT are indicated here, and for DeepSeek — here.
The Chinese managed to save money partly thanks to the "mixture of experts" approach. This is a method of working with information that involves separate processing of data, where a corresponding expert approach is applied for each specific task or area. This scheme helped developers save time on training the model and organize work on a cluster of relatively inexpensive NVIDIA H800 graphics cards. According to the project team, they spent about 20 times less on creating and maintaining their language model than their competitors.
❗ Additionally, the "Chinese" has open-source code. This means that anyone can customize the neural network to their requirements. Users of ChatGPT could only dream of such an opportunity.
How Much Does DeepSeek Cost?
Here are the rates listed on the official website of the company:
Here’s an explanation of how to calculate the cost (spoiler: it's complicated):
1. Calculating the Number of Tokens
- Input Tokens: These are the data you send to the model.
- Output Tokens: This is the text generated by the model.
- For the deepseek-reasoner model, you need to account for both tokens used in the chain of thought (CoT) and tokens for the final answer.
2. Price per Token
- The price depends on the model and the number of tokens.
- For example, if the price for 1 million tokens is $0.14, then for 10,000 tokens (e.g., 5,000 input and 5,000 output), the cost would be: 10,000 tokens × ($0.14 / 1,000,000 tokens) = $0.0014
3. Discounts
- Discounts are available until February 8, 2025, for all models except DeepSeek-R1. After that date, prices will return to normal.
4. How Payments Are Processed
- Funds are deducted from your balance (first from any provided balance if available). The amount deducted depends on the number of tokens used.
Simple Cost Calculation
Number of Tokens × Price per Token = Cost.
Suppose we need to generate an article through the Chinese language model that ultimately contains 10,000 words. Here’s how the calculations would look:
To calculate the cost, we need to understand how many tokens will be required for 10,000 words.
1. How Many Tokens Are in 10,000 Words?
On average, 1 word is about 1.3 tokens. Therefore, 10,000 words would be approximately 13,000 tokens.
2. Price for Tokens
Assuming the price for 1 million tokens is $0.14 (as in the example).
3. Cost Calculation
If 10,000 words equal 13,000 tokens, then the cost would be:
13,000 tokens × (0.14 /1 000 000) = $0.00182.
Thus, generating text of 10,000 words would cost approximately $0.00182. This is an example, and the exact price depends on the model used for text generation.
❗ Interesting fact: in order not to lose users thanks to the triumph of the Chinese alternative, the OpenAI team urgently introduced promo codes and started giving away other discounts.
ChatGPT calculates payment in a different way
You have to pay for the chatbot according to the tariffs: monthly or per year. There are different tariff plans, depending on your needs.
It is noteworthy that ChatGPT itself (version o1) admits that its competitor offers much more favorable working conditions.
Conflict Between ChatGPT and DeepSeek
The OpenAI team claims that DeepSeek used their model to train its chatbot. This statement raised concerns about potential intellectual property violations.
OpenAI asserts that it has found signs of "distillation" techniques being used — a method by which developers can enhance the performance of smaller models using insights from larger, more powerful models. This allows for achieving similar results in specific tasks at much lower costs.
OpenAI has not disclosed details of its evidence but indicated that its terms of service state that users cannot "copy" its services or use their outputs to develop competing models.
Users on social media found OpenAI's claims amusing.
"Sorry, I can't stop laughing. OpenAI, a company built on the actual theft of the entire internet, is crying because DeepSeek may have trained on outputs from ChatGPT," writes journalist Ed Zitron, reminding everyone of OpenAI's controversial training methods.
There have also been reports circulating that DeepSeek initially presented itself as ChatGPT. According to those posting about it, developers quickly fixed this bug. The conflict between ChatGPT and DeepSeek quickly became a source of memes.
"OpenAI to DeepSeek: You're trying to steal what I've already stolen."
There were also jokes about the reaction of ChatGPT developers to the emergence of a worthy competitor.
Users were particularly amused by the speed with which DeepSeek entered the "battle" between ChatGPT and Google's Gemini.
The teams of these neural networks had been competing for the title of market leader for a long time. However, with the emergence of DeepSeek, something went awry.
DeepSeek Under Fire
Many view the confrontation between DeepSeek and ChatGPT as yet another political battle between the USA and China. Reports have surfaced about a potential ban on the Chinese neural network in America. Meanwhile, the media is reporting that lawmakers are urging Trump to consider new restrictions on Nvidia chips used by DeepSeek.
According to Bloomberg, just a few days after the release of the latest model of Chinese artificial intelligence, efforts began in hundreds of companies to impose a ban on it. Additionally, there are reports in the media about an investigation by Microsoft and OpenAI against DeepSeek.
In Italy, the Chinese neural network has already been banned. The country's data protection authority prohibited DeepSeek due to a lack of information on how it uses personal data. The neural network is also under scrutiny from French and Irish regulators.
What is Better: ChatGPT or DeepSeek in Reasoning Capabilities?
Opinions are divided on whether ChatGPT or DeepSeek is better. Some assert that the Chinese neural network far surpasses its American competitor, while others believe that no matter how expensive ChatGPT is, DeepSeek still has a long way to go.
Professor Mushtaq Bilal compared the effectiveness of both companies' models. He specializes in the ethical use of artificial intelligence for academic purposes. The professor presented his results in a graph.
Here’s how to read it:
The graph shows accuracy results and percentiles (a statistical measure that divides data into 100 equal parts) across several tests for AI models: DeepSeek-R1, OpenAI-o1-1217, DeepSeek-R1-32B, OpenAI-o1-mini, and DeepSeek-V3. Here are the names of the tests:
- AIME 2024 (Pass@1): A test measuring the percentage of correct answers.
- Codeforces (Percentile): Evaluation based on results from competitions on the Codeforces platform.
- GPQA Diamond (Pass@1): A test for solving questions on the first attempt
- MATH-500 (Pass@1): A test focused on solving mathematical problems.
- MMLU (Pass@1): A multitasking test
- SWE-bench Verified (Resolved): A test for solving software development tasks.
Judging by the test results, DeepSeek-R1 outperforms other models in terms of correct answers. In the Codeforces test, ChatGPT 01 took first place. It also wins against its competitor in solving questions on the first attempt. However, when it comes to solving mathematical problems, DeepSeek has no equal. The Chinese neural network also surpasses its rival in programming problem-solving tasks. On the other hand, ChatGPT excels in multitasking scenarios.
Here’s what else users are saying about DeepSeek:
| What Users Criticize DeepSeek For | Censorship. For example, the chatbot does not answer questions about Chinese politics. | No filters for 18+ content. | The neural network may forget the conditions of the prompt. | The neural network lures users into a paid subscription. |
| What Users Praise DeepSeek For | Performs well with translations. | Offers free access. | Users claim that the Chinese chatbot passes AI detector tests better than ChatGPT. | Works without a VPN. |
The buzz around DeepSeek had barely settled when news emerged that the Chinese neural network now has a competitor — the Qwen artificial intelligence from Alibaba. Company representatives, in the best marketing traditions, claim that their product is more effective than its counterparts. However, there are also other Chinese language models that could join the race in the future.
Instead of Conclusions
With the emergence of DeepSeek, users have begun to talk about the start of a new era for language models. In this race, developers will try to outdo their competitors by any means necessary. OpenAI's reaction to the arrival of a Chinese alternative is a vivid testament to this.Tests show that DeepSeek already surpasses ChatGPT in several metrics. However, the Chinese neural network also has its drawbacks. Nevertheless, its accessibility outweighs some of these shortcomings.
