We’ve all grown accustomed to ChatGPT and actively use it for work or everyday tasks. This model has become one of the fastest-growing technologies in human history. Within just two months of its launch in November 2022, ChatGPT reached 100 million users (for comparison, Google Gemini, introduced in March 2023, had around 42 million active users by October 2024).
However, a new powerful player has entered the scene – the Chinese model DeepSeek. This Chinese AI app became the most downloaded free app in the U.S. App Store, outperforming not only ChatGPT but also all the top social networks, marketplaces, and other tech giants.
In Google Play, the situation is not much different – the AI developed by a team of just 200 people has already surpassed one million installs, and the app itself is less than two weeks old.
In this article, we want to delve into why DeepSeek has become such a hot topic in recent days. Let’s explore which tool is better suited for your tasks and whether DeepSeek has the potential to replace OpenAI’s product in the market.
Global Reactions to the News
On January 27, 2025, the shares of leading global tech companies specializing in chips, graphics cards, and computational equipment experienced a sharp decline. Nvidia lost 17% of its market capitalization, Dell's stock dropped by 7.2%, and the Nasdaq index fell by 3.04%. Other giants were also affected: Microsoft shares fell by 2.2%, and Alphabet lost 3.5%. The main reason behind this downturn was the launch of the large language model (LLM) R1 by DeepSeek, positioned as a direct competitor to OpenAI's o1.
Why Did DeepSeek's Launch Cause a Global Stock Market Drop?
The company claims that their DeepSeek R1 model was developed in just two months at a cost of only $6 million. Allegedly, they utilized the computing power of Nvidia H800 chips, which are subject to U.S. sanctions against China. Despite the fact that the H800 is not Nvidia's most advanced technology, DeepSeek managed to create a model capable of competing with the most powerful AI models developed in the U.S.
If DeepSeek's claims prove accurate, it could challenge the necessity of massive investments in AI infrastructure. Their approach demonstrates that it’s possible to develop powerful solutions "quickly and efficiently." The cost of processing DeepSeek-R1 requests is just $0.55 per million tokens, compared to $15 for OpenAI.
For those who are not familiar: what do these metrics mean?
- Training approach – DeepSeek R1 uses minimal human supervision, while OpenAI o1 undergoes extensive fine-tuning with human feedback (RLHF).
- Input token cost – The cost of processing user input. DeepSeek charges $0.14–$0.55 per million tokens, while OpenAI costs $7.50–$15.
- Output token cost – The cost of generating responses. DeepSeek charges $2.19, while OpenAI is $60 (27 times more expensive).
- Codeforces benchmark – A test for programming-related tasks. The results are almost identical: 96.3% (DeepSeek) vs. 96.6% (OpenAI).
- Math benchmark – Measures performance on math problems. DeepSeek scored 79.8/97.3, while OpenAI scored 79.2/96.4.
DeepSeek is significantly cheaper, while its performance in coding and math is nearly identical to OpenAI o1.
The success of DeepSeek calls into question the need for enormous expenditures on AI development. Their example shows that competitive models can be built both quickly and cost-effectively.
ChatGPT offers free versions, which is suitable for basic tasks but comes with limitations. For $20/month, ChatGPT Plus provides access to the more powerful GPT-o1 with limits 30 requests per week, while the Pro version at $200/month offers unlimited use of the o1 model. Meanwhile, DeepSeek is currently free, making it a cost-effective option, although paid plans may be introduced in the future (though there is no official information on this yet).
Although ChatGPT’s Plus subscription is not free, it offers many additional features that make it a more versatile tool. One of its key advantages is advanced voice mode. Another useful feature is image generation. However, it’s worth acknowledging that the image quality is not the best when compared to other competitors on the market.
For developers, ChatGPT provides live code editing directly in the chat (Canvas mode), as well as integration with IDEs. It also has a dedicated desktop application, making it more convenient to use on a computer.
Additionally, ChatGPT’s web search is more advanced, delivering faster and more accurate results compared to DeepSeek. These features make it a strong choice for those who need more than just basic AI text generation.
Let’s Get Practical
But let’s move away from the dry numbers and focus on what truly matters to users. After all, the most popular use cases for artificial intelligence remain text and code generation. So, why not put these models to the test and see what they can really offer? Let’s dive into practical examples and compare their performance in tasks that genuinely interest everyday users
It is worth clarifying that during the experiment, ChatGPT o1 (the most powerful model available at a reasonable subscription price) and DeepSeek R1 were used.
Who Writes Better?
We decided to start by testing text generation. We gave both models a detailed and clear prompt to ensure a fair comparison:
Create a promotional text for a social media post advertising a local coffee shop. The text should be short, engaging, and include a call to action.
Instructions:
A local coffee shop is looking for ways to attract new customers. Their key selling points are freshly roasted coffee and in-house baked pastries.
- Target audience: Young people and office workers in the city center who are looking for high-quality coffee and a cozy place to work or meet.
- Requirements: The text should be no longer than 50 words and must include a call to action.
Of course, this LLM model handled the task brilliantly. All the requirements were strictly followed, and even the word count was precisely 43, which impressed me — previous models often struggled to adhere to character limits. However, we have to admit that the text feels somewhat formulaic. For example, the phrase "...work, study, or a quick escape..." is well-structured, but it sounds a bit mechanical. So, let’s see how its competitor performed.
The first thing worth noting is that DeepSeek demonstrates its logical reasoning process, explaining the steps it went through before arriving at an answer. We may still be far from achieving general artificial intelligence, but observing this level of "thinking" in a machine makes you realize just how far technology has already come.
So, here’s my feedback on the result. Personally, we prefer this version (though we can’t ignore the fact that ChatGPT’s writing style might simply feel overused at this point). The response included hashtags, which is great since the text is meant for social media. It also provided an exact word count (a nice touch, considering we specifically requested a text under 50 words) and clearly outlined the CTA. Additionally, there was space left for the café’s name, whereas the competitor used a random placeholder. We wouldn’t say the text feels entirely human-written, but overall, we like the structure and presentation more.
Summing up everything mentioned above, in this round, we believe DeepSeek has the advantage. It approached the task more thoughtfully, even though ChatGPT’s text was still high-quality and met all the given requirements.
Who Handles Code Better?
We decided to test how these two models perform in writing and optimizing code. The task was simple yet revealing: there is a function that runs in 1.1 seconds (1100 ms), and the goal is to make it faster.
private function getDisabledSpecValueIds(Collection $specValues, array $selected, array $relations): array
{
$callback = function () use ($specValues, $selected, $relations) {
$disabledIds = [];
foreach ($specValues as $specsWithValue) {
foreach ($specsWithValue->specValues as $value) {
// Add current value
$curSelected = $selected;
$curSelected[$specsWithValue->id] = [$value->id];
$disabled = true;
// All products
foreach ($relations as $relation) {
$goodProduct = true;
// All selected specifications
foreach ($curSelected as $specId => $valuesIds) {
if (!isset($relation[$specId])) {
$goodProduct = false;
break;
}
foreach ($valuesIds as $vId) {
if (in_array($vId, $relation[$specId])) {
continue 2;
}
}
$goodProduct = false;
break;
}
if ($goodProduct) {
$disabled = false;
break;
}
}
if ($disabled) {
$disabledIds[] = $value->id;
}
}
}
return $disabledIds;
};
// Cache result
$cacheKey = sha1(json_encode([
'disabledSpecValueIds',
'specificationValueIds' => $specValues->pluck('specValues.id')->sort()->all(),
'selected' => $selected,
'relations' => $relations,
]));
if (Cache::supportsTags()) {
return Cache::tags(ProductFilterSpecificationValuePivot::getCacheTagName())
->rememberForever($cacheKey, $callback);
}
return Cache::remember($cacheKey, now()->addHour(), $callback);
}
DeepSeek R1 immediately identified the bottleneck in the code and suggested a solution.
Most interestingly, after applying the changes, the execution time decreased by 25 times (from 1100 ms to 40 ms).
Let's continue our testing. ChatGPT o1 also identified problem areas and suggested its own solution.
I’m not sure why, but ChatGPT o1 didn’t fully understand how to retrieve data from relations and provided only a schematic suggestion. We had to modify the buildIndex function to match the approach proposed by R1.
However, the existsAnyProductForCombination function from o1 turned out to be twice as effective as the one suggested by R1. As a result, o1’s solution now runs in 20-22 ms, though it required a slight adjustment.
DeepSeek R1 excels at analyzing code and provides clear, easy-to-implement solutions. However, while ChatGPT o1 required a bit more intervention, it ultimately delivered a more efficient solution. So, in this round, the victory goes to ChatGPT — it managed to extract the maximum speed from the code.
Bottom Line
The rise of DeepSeek shows that the era of AI being dominated by just a few major players is coming to an end. Now, even relatively small teams can develop competitive models by focusing on smart optimization rather than massive computational resources. This shift could not only reshape the AI market but also change the way such technologies are developed.
Which model is better? It depends on your needs. If you prioritize speed and cost efficiency, DeepSeek has a clear advantage. If you need additional features like voice mode, image generation, or code integration, ChatGPT might be the better choice.
My advice — try both and see which one works best for you. A new and even more powerful player may emerge sooner than we think.
