I know you've seen it. It's been dominating the timelines since the end of January 2025 and you're only bound to hear more about it as the year progresses. China once again takes the world by storm with its new contribution to the international zeitgeist. And no, I'm not talking about another Covid outbreak (Thank God).  Responsible for sinking US stocks and sending the American tech community in a panic, let's see why DeepSeek has the entire world in a frenzy. 

What exactly is DeepSeek? 

DeepSeek is a Large Language Model,  similar to the renowned ChatGPT. It entered the artificial intelligence arena in late 2023 and was created by former ByteDance (the creators of TikTok) AI researchers. In simple terms, it can process enormous amounts of data and create high-quality human-like outputs. 

Although it's up for debate, DeepSeek purportedly works just as well as OpenAI's latest model. Its claim to fame however, is boasting lower memory usage and costing a fraction of the mount of money to train compared to ChatGPT (we are talking a difference of $100 million to $6 million dollars). 

The New Cold War: ChatGPT vs DeepSeek

The United States finds itself in another technical war with the superpower. nVidia, considered the undisputed winner of the AI and high-performance commuting race, lost $600 million in valuation once the newest model of DeepSeek was released. 

Now comes the questions everyone wants to know the answer to: "What are the differences between the two?" and "Which one is better?" 

Well, we're here to give you the quick rundown.

 Mathematical Reasoning

Let's keep things simple. The MATH-500 is a mathematics benchmark containing 500 problems across a variety of mathematical subjects. As you could've guessed, higher scores indicate higher mathematical problem-solving abilities.

Deepseek-R1 scored 97.3%. 

OpenAI o1 (ChatGPT's counterpart for advanced reasoning tasks; both developed by OpenAI) scored 96.4%

Both are extremely capable at advanced mathematics but the slight margin suggests that DeepSeek is slightly more capable. 

Coding

Pretty self explanatory. For this metric, something called Codeforces was used to measure the coding capacity of each AI model. It's a website that hosts competitive programming competitions and has its own rating system. 

DeepSeek-R1 scored 2029.

OpenAI o1 scored 2061. 

Just so you know, the 1900 to 2099 score is the boundary for Candidate Master (right below Master) so it's safe to say both these systems are more than proficient at coding, with OpenA1 still maintaining an edge. 

General Reasoning

This is judged via GPQA (Graduate-Level Google-Proof Q&A) Diamond benchmark assessments. Diamond being the most difficult level of these kind of assessments. They judge the capabilities of AI models with difficult questions across a variety of scientific topics. Let's see how each model performed.

DeepSeek-R1 scored 71.5%.

OpenAI o1 scored 75.7%.

Once again, while relatively close, OpenAI outperforms DeepSeek in this metric.

Cost Comparison

Now here is where the rubber meets the road. While OpenAI has slightly outperformed DeepSeek in two out of the three categories listed above, we know that money talks and everything else walks. Let's look at the financial efficiency of each of these models which, considering how closely they perform, will not doubt have the greatest effect on their adoption in the market going forward. 

This can be broken down into price per tokens (typically calculated by units of one million, or 1M). Tokens represent, for all intents and purposes, words (or pieces of words, i.e. characters). These tokens have a real world price equivalent to be input, processed, and output, respectively. That's the gist. Now, how much does it cost for either to process this information?

At the time of this writing, DeepSeek costs about $0.14 per 1 million tokens for input. The cost to cache 1 million of these input tokens is $0.55. The price output of 1 million tokens is $2.19.

OpenAI o1, however, costs $7.50 per 1 million tokens for input, $15.00 per 1 million tokens for these inputs to be cached, and a hefty $60.00 per 1 million tokens for output. 

You can see the dramatic difference in price for near identical performance metrics, making DeepSeek the cost effective option by far for large-scale AI implementation.

Its still relatively early in this competition to determine a definite victor. Depending on who you ask, there are many detractors and advocates for both, but one thing we can say for sure. The financial, technological and human capital allocated toward the artificial intelligence race will rival the Space Race of the 50's and 60's.

Citations