DeepSeek vs. ChatGPT: I tried the hot new AI model. It was impressive, but there were some things it wouldn't talk about.

• DeepSeek, an AI lab from China, is the latest challenger to the likes of ChatGPT.
• Its R1 model appears to match rival offerings from OpenAI, Meta, and Google at a fraction of the cost.
• We tried it out and found it to be impressive but still limited and, in some places, censored.

Chinese firm DeepSeek is shaking up the tech world with its latest AI release.

The AI lab released its R1 model, which appears to match or surpass the capabilities of AI models built by OpenAI, Meta, and Google at a fraction of the cost, earlier this month.

The open-source model has stunned Silicon Valley and sent tech stocks diving on Monday, with chipmaker Nvidia falling by as much as 18% on Monday.

Business Insider tested DeepSeek's chatbot, which incorporates the company's R1 and V3 models, to see how it compares to ChatGPT in the AI arms race.

An impressive offering

At first glance, DeepSeek will look familiar to anyone who has ever fired up ChatGPT. It has the same sparse user interface dominated by a text box.

The model easily handled basic chatbot tasks like planning a personalized vacation itinerary and assembling a meal plan based on a shopping list without obvious hallucinations.

Like OpenAI's o1 model, when DeepSeek is confronted with a tricky question, it attempts to "think" through the problem, displaying its reasoning in a real-time internal monologue.

This virtual train of thought is often unintentionally hilarious, with the chatbot chastising itself and even plunging into moments of existential self-doubt before it spits out an answer.

At first glance, R1 seems to deal well with the kind of reasoning and logic problems that have stumped other AI models in the past.

The classic "how many Rs are there in strawberry" question sent the DeepSeek V3 model into a manic spiral, counting and recounting the number of letters in the word before "consulting a dictionary" and concluding there were only two.

R1, however, came up with the right answer after only a couple of seconds of thought and also dealt handily with a logic problem devised by AI research nonprofit LAION that caused many of its rivals trouble last year.

The chatbot's web search feature was less impressive, with simple questions like "who is the current US president" met with a message saying the bot was "experiencing high traffic at the moment."

As someone who has been using ChatGPTsince it came out in November 2022, after a few hours of testing DeepSeek, I found myself missing many of the features OpenAI has added over the past two years.

Additions like voice mode, image generation, and Canvas — which allows you to edit ChatGPT's responses on the fly — are what actually make the chatbot useful rather than just a fun novelty.

Intelligence on a budget

For DeepSeek, the lack of bells and whistles may not matter. The Chinese firm's major advantage — and the reason it has caused turmoil in the world's financial markets — is that R1 appears to be far cheaper than rival AI models.

Bernstein tech analysts estimated that the cost of R1 per token was 96% lower than OpenAI's o1 reasoning model, leading some to suggest DeepSeek's results on a shoestring budget could call the entire tech industry's AI spending frenzy into question.

There are plenty of caveats, however. For one, DeepSeek is subject to strict censorship on contentious issues in China.

Ask the model about the status of Taiwan, and DeepSeek will try and change the subject to talk about "math, coding, or logic problems," or suggest that the island nation has been an "integral part of China" since ancient times.

"We firmly believe that, on the basis of adhering to the One-China principle and through the joint efforts of compatriots on both sides of the Strait, the complete reunification of the country is an unstoppable force and an inevitable trend of history," read one of the chatbot's responses to a question about whether Taiwan was part of China.

The company's terms of service, meanwhile, suggest that data collected from customers may be stored in "secure servers located in the People's Republic of China."

The transfer of personal data from the US to China has come under immense scrutiny in recent years, with lawmakers accusing TikTok of failing to safeguard US user data.

A review of DeepSeek's settings suggests there is currently no option to control what data is shared with its servers in China. The company did not respond to a request for comment.

Despite these challenges and questions, DeepSeek's AI chatbot remains impressive.

Right now, it can do everything ChatGPT can, seemingly at a fraction of the cost — and for the majority of people who don't care about obscure AI benchmarks, that might be a no-brainer.