DeepSeek vs ChatGPT

DeepSeek has received a lot of attention lately after launching its new R1 model, said to be on a level with ChatGPT. The launch created a storm in the market, and large companies such as Nvidia and OpenAI dropped on the stock market. The DeepSeek iOS app is now also the most downloaded free app in the USA, having overtaken the OpenAI app at the top.

Is DeepSeek-R1 really better than ChatGPT? And not least, is it better at coding?

Nvidia has dominated the market for artificial intelligence chips and controls roughly 80% of it. The company sells the most expensive AI chips, or graphics processors (GPUs), in the world. For Nvidia to continue its success it has to remain the preferred choice for the AI needs of businesses, and customers have to keep investing in its newest and most expensive chips.

But at the end of January something happened that raised questions about the ability of Nvidia to sustain that strong revenue growth over the long term. DeepSeek announced that it had trained its R1 model in just two months for under 6 million dollars. By comparison, the largest American technology companies have spent billions of dollars on Nvidia chips. As a result the Nvidia share fell 17% in a single trading day, equivalent to a loss in value of almost 600 billion dollars.

DeepSeek is a Chinese AI company founded in 2023 that builds open-source large language models (LLMs). If you are new to large language models, this is a type of machine learning model trained on a very large amount of data in text form. The result is a model that generates or predicts human-like text.

DeepSeek-R1 is an AI model developed by DeepSeek and launched in January 2025. The model competes with, and in some cases surpasses, the reasoning ability of some of the most advanced models in the world, but at a fraction of the running cost, according to the company itself. DeepSeek-R1 is also open source under the MIT licence, which allows free commercial and academic use.

The fact that DeepSeek-R1 is open source means the public has free access to both the model and the source code, which can have large implications for AI development going forward:

- Lower development costs. Nvidia chips have been expensive and a large cost item in AI development. With an open model like DeepSeek-R1, developers can train and use advanced AI without being as dependent on the Nvidia ecosystem.

- Adaptation and further development. The source code and the model are open, so developers can fine-tune the model for specific use cases, train it further with their own data or optimise it for their own needs. Developers can run the model on their own servers or in the cloud without depending on a third-party supplier, which gives better control over data, performance and costs.

- Transparency. Since the code and architecture are open, developers can understand how the model works, evaluate its strengths and weaknesses, and identify potential bias or security problems.

- Integration with your own systems. Open access to the model makes it easier to integrate it into existing solutions, APIs or applications without the limitations of commercial APIs.

This will be particularly important for Europe, since it offers a way to join the AI race. Europe has long depended on AI models from American players such as OpenAI, Google DeepMind and Meta. Open-source models give European companies and researchers the chance to build their own solutions without being locked to American cloud services.

By being able to use, adapt and host the model freely themselves, European players can secure data sovereignty, comply with strict privacy requirements such as the GDPR, and develop tailored solutions for multilingual and industry-specific needs. At the same time the DeepSeek API gives considerably cheaper access to advanced AI, making it possible for European start-ups and companies to compete globally without the enormous costs often tied to AI development.

One of the biggest differences between DeepSeek-R1 and ChatGPT is focus. ChatGPT is primarily designed for conversation and focuses on narrow AI, task-specific intelligence. DeepSeek, on the other hand, is trying to reach artificial general intelligence (AGI).

Artificial general intelligence is a hypothetical form of AI that can understand and learn any intellectual task the way a human can. The goal is to mimic the cognitive abilities of the human brain. The new model is built to solve complex intellectual tasks at a level approaching human capacity, with a particular focus on improved reasoning. That makes it an important contribution to the ongoing development of more powerful and flexible AI systems.

Earlier, OpenAI set a standard for reasoning-based AI with its o1 model, which uses chain-of-thought techniques to break problems into several steps. Using reinforcement learning, the model can improve its strategies by adjusting based on reward systems, which helps it identify errors and test alternative solutions when needed.

DeepSeek-R1 builds further on that approach by combining reinforcement learning with supervised fine-tuning, which makes the model especially well suited to demanding logical and mathematical tasks. The test results show strong performance, with 79.8% on the AIME 2024 mathematics test, 97.3% on the MATH-500 benchmark, and a Codeforces rating of 2,029, where it surpasses 96.3% of human programmers. By comparison, the OpenAI o1-1217 achieved 79.2% on AIME, 96.4% on MATH-500 and 96.6% on Codeforces.

On general knowledge, DeepSeek-R1 performed with 90.8% accuracy on the MMLU benchmark, just one percentage point behind the 91.8% of o1. These results underline that DeepSeek-R1 is one of the most advanced open AI models on the market and demonstrate progress in reasoning and problem solving towards AGI.

There is also a difference in the architecture of the models. DeepSeek-R1 uses a Mixture-of-Experts (MoE) approach, which can be compared to a team of specialised experts where only the most relevant are activated for each task. With 671 billion parameters in total, DeepSeek-R1 activates only a subset of them (37 billion) for each request, which increases efficiency. This MoE architecture lets DeepSeek-R1 optimise both performance and resource use by adapting dynamically to different types of question. By comparison, ChatGPT uses a traditional transformer model where all parameters are activated for every task, which gives consistent results but potentially lower efficiency.

Each model has its strengths in different areas. DeepSeek-R1 stands out particularly in technical tasks, with impressive results in mathematics for example. That makes it especially valuable for tasks requiring precise technical solutions. ChatGPT, on the other hand, has a stronger ability to understand context and deliver more nuanced answers across a broader range of subjects.

- Writing. Both models can help with documentation and content production, but their approaches differ. ChatGPT is skilled at creating engaging, conversational content with broad context, which makes it ideal for explaining complex data concepts to non-technical stakeholders. DeepSeek, by contrast, stands out in technical writing tasks, producing precise and formal documentation. That makes it especially valuable for technical specifications and documentation of data-related projects.

- Creativity and proposed solutions. ChatGPT is excellent at suggesting several varied solutions, giving a broad range of analytical options to explore. DeepSeek tends to focus on fewer but more thoroughly worked solutions, going deep on one approach. That makes it especially useful when you need a detailed, carefully designed data strategy.

- Research and learning. ChatGPT offers more tutorial-style explanations that work well for learning new concepts. It is particularly good at breaking complex topics into understandable parts. DeepSeek prioritises precision and brevity, which makes it useful for quick lookups and fact checking, for example. It is particularly valuable when investigating specific methods or algorithms, or other technical questions.

For programmers, choosing the right AI tool matters. ChatGPT offers extensive coding help with detailed explanations, which makes it an excellent learning tool. DeepSeek-R1 focuses on faster code generation and more precise answers, which is especially useful for quick, efficient solutions to specific programming challenges.

ChatGPT probably still gives better answers when you want to understand complex implementations, but DeepSeek manages to solve the harder tasks more precisely because it puts more focus on one solution rather than an expansive answer with lots of context and explanation. ChatGPT tends to offer several different suggestions for a problem, which can be both positive and negative depending on the situation.

DeepSeek-R1 is still relatively new and has both advantages and drawbacks to its approach. During testing we found the servers very often unavailable, which has been a drawback. With a little maturing we think DeepSeek can become a real competitor to GPT.

DeepSeek is still free and open source, which can give developers the chance to build their own tailored models. API access is also much cheaper than with its competitor OpenAI.

DeepSeek R1 is specifically designed for technical and structured tasks. It handles coding, mathematical reasoning and logic-based questions with high efficiency, which makes it a good choice for many developers and researchers. If you are looking for precise, direct answers without unnecessary detail, DeepSeek R1 delivers that efficiently.

ChatGPT o1, on the other hand, is more flexible. It performs well in creative writing, brainstorming and open discussion, which makes it ideal for content production, research and informal conversation. Although it can handle technical subjects, it tends to give more detailed explanations, which can be useful for users who prefer more context.