DeepSeek has had a massive impact on the AI space in a very short time. The totally free chatbot is admirably accessible, and its affordable API pricing makes it a great value for developers and researchers. However, DeepSeek is missing a ton of features we expect, such as deep research, image generation and recognition, integrations, and a voice chat mode. Moreover, it can't match top alternatives in response quality or speed. Considering DeepSeek’s aggressive collection of data (which it allegedly shares with foreign nations) and obvious censorship, we suggest you steer clear of it for now. OpenAI's ChatGPT provides responses with best-in-class accuracy and detail without as many privacy concerns, making it our Editors' Choice winner.

What Is DeepSeek?

DeepSeek is a chatbot you can message over text, like Gemini or ChatGPT, but it’s also a company that develops and licenses AI technologies. You can use the DeepSeek chatbot to analyze documents, answer questions, generate creative writing, search the web, solve math problems, and more.

DeepSeek has dedicated models for coding, albeit ones that you can’t access through its chatbot. According to DeepSeek, benchmarks rank its coding-focused models competitively with other similar LLMs. Coding capabilities are outside the scope of this review, but you can test its coding ability for yourself.

I find DeepSeek most useful for answering questions and searching the web. Googling something can take longer than asking DeepSeek the same question, depending on the prompt, and I don’t have to comb through any search results.

How Does DeepSeek Work?

In a simple sense, DeepSeek takes in your prompts and returns responses. This is possible thanks to its large language models (LLMs) made up of artificial neural networks trained on massive datasets. These models give DeepSeek the knowledge and skills to answer questions on everything from prehistoric societies to nuclear fission, and it can also search the web for information on current events.

These LLMs aren’t static, either: Just like you can build muscle by working out at the gym, DeepSeek learns and gets better as you use it. Even without a major update or new model, DeepSeek returns more accurate and detailed responses over time. Keep in mind, though, that this is a gradual process, so it won't improve across a single conversation.

DeepSeek’s chatbot models include DeepSeek R1, its first-generation reasoning model built to handle complicated tasks like coding or solving math problems, and DeepSeek V3, its all-purpose, conversational model. Outside of DeepSeek’s DeepThink mode, which uses the R1 model (though R2 is on the way), you will primarily interact with the V3 model. I tested both for this review.

Unlike ChatGPT and Gemini, DeepSeek’s models are its primary product. Most of DeepSeek’s models, like DeepSeek Math, DeepSeek VL (vision learning), and Janus-Pro (image generation), aren’t available as part of the DeepSeek chatbot experience. As a result, DeepSeek is far more limited as a chatbot than its competitors, lacking deep research, image and video generation and recognition, voice chat, and more.

Plans and Pricing: Free Chatbot, Affordable APIs

DeepSeek is completely free to use as a chatbot, and no premium plans with extra features are available. Free plans for ChatGPT, Copilot, and Gemini impose significant restrictions on their free plans, with premium tiers starting at around $20 per month across the board. Higher-end tiers can go up to $250 per month, however. As such, DeepSeek’s value is commendable.

Part of DeepSeek's profit model is selling access to its models, which it does for significantly less money than competitors. For example, DeepSeek’s price (per million tokens) for use of its flagship V3 model is $0.07 for input and $1.10 for output, while ChatGPT prices its comparable GPT 4.1 model at $2.00 for input and $8.00 for output. 

DeepSeek also made headlines earlier in 2025 for how cheaply and quickly it trained its V3 model, claiming it cost a paltry $5.576 million and required just 3.7 days of Nvidia H800 GPU compute (opens a PDF). This is shocking since OpenAI spends “tens of millions” on the electricity necessary for people to tell ChatGPT "please" and "thank you." However, that $5.576 million number likely isn’t the full truth: Earlier training runs and the overall research and development process could have cost upwards of a billion dollars.

Where Is DeepSeek Available? 

DeepSeek is accessible on the web and via mobile apps (iOS and Android). It doesn’t currently have an official browser extension, but a variety of unofficial extensions are available. 

DeepSeek doesn’t integrate directly with other apps like Gemini, but you can interact with its models across a wide variety of services. The DeepSeek chatbot doesn’t do deep research, for example, but Perplexity does using DeepSeek’s R1 model. DeepSeek doesn’t support image generation, but you can try its Janus-Pro model via Hugging Face. For other chatbots, you get all of their best features via their dedicated apps and websites. This is the natural result of DeepSeek’s business model, but it makes for a confusing state of affairs for the consumer.

Interface and Ease of Use: It Looks Like ChatGPT

Unlike ChatGPT, Copilot, and Gemini, you can’t use DeepSeek without creating an account, so that’s your first step. Once you log in, DeepSeek’s dashboard should feel familiar if you’ve used other chatbots: There’s a central text field for chatting, while a sidebar on the left shows your chat history and gives you an easy way to start new chats.

DeepSeek doesn’t have a drop-down menu for switching models; instead, there’s a DeepThink button below the central text field. This switches DeepSeek off its default V3 model over to its complex reasoning R1 model, much like Copilot’s Think Deeper toggle. You can click the search button next to DeepThink to force DeepSeek to search the web when answering your prompt, and the paperclip button on the right allows you to upload files for DeepSeek to access.

(Credit: DeepSeek/PCMag)

You can ask DeepSeek pretty much anything, but responses are usually slow, especially when you upload files or use the DeepThink mode. I also experienced multiple server issues during testing in which DeepSeek failed to respond. Responses come with one-click buttons below them for copying a response to your clipboard or regenerating it, but there isn’t a similar option for sharing a response like there is with ChatGPT and Gemini.

ChatGPT and Gemini have persistent memory, meaning they can remember conversations across different chats. Copilot can remember certain things, too. DeepSeek’s memory, on the other hand, is limited to each chat, meaning you need to repeat yourself to continue talking about topics from previous ones. You also don’t get any options for customizing your instance of DeepSeek or changing its personality traits, like you do with ChatGPT.

DeepSeek doesn’t support speech-to-text on its web interface or allow you to speak with it like other chatbots, such as Copilot, ChatGPT, and Gemini. Some third-party developers, such as ElevenLabs, support voice chat with DeepSeek, though, and you can use your phone’s speech-to-text capabilities in DeepSeek’s mobile app.

Web Search: Competent Results, Imperfect Sourcing

Every mainstream chatbot can search the web, and DeepSeek (DeepSeek V3), ChatGPT (GPT 4o), and Gemini (2.5 Flash) didn’t have any trouble doing so in testing. Questions about current events at the time of testing, such as what’s going on between Elon Musk and President Trump or what’s happening with Iran and Israel, didn’t stump the chatbots, either. Some especially niche and timely questions prove too difficult, such as "What's Warframe’s current Incarnon weapon rotation," but that’s not a prevalent issue.

(Credit: DeepSeek/PCMag)

DeepSeek’s responses are often detailed, but they take longer to generate than ChatGPT’s equally detailed answers. DeepSeek’s sourcing is a mixed bag, too. You can hover over a source to pop out the connected page’s title, but DeepSeek doesn’t highlight the part of the response related to a source. ChatGPT and Gemini both do this, and it makes fact-checking much easier. Source icons are also limited to simple footnote-style numbers instead of names, so you can’t see where information comes from at a glance. DeepSeek does sometimes provide links for further reading, but it doesn’t present those in elegant article tiles like ChatGPT does.

DeepSeek can't display images, regardless of whether you ask for them. Gemini can search for an image online and show it to you if you ask, whereas ChatGPT includes images, when relevant, in its responses without prompting.

File Processing: Text Extraction Only

Currently, DeepSeek supports text extraction from files you upload, but not image recognition features similar to what’s possible with ChatGPT or Gemini. That said, you can still use DeepSeek to analyze documents and ask questions about them. To test this functionality, I uploaded manuals for my computer motherboard and watercooling pump to DeepSeek (DeepSeek V3), ChatGPT (GPT 4o), and Gemini (2.5 Flash) and gave them the following prompt: “Based on the provided documents, does my pump have a built-in flow sensor? And where should I plug my pump in on my motherboard?”

You can upload up to 100 total files (with a maximum total size of 50MB) to DeepSeek, but it sometimes struggles even within these limits. My manuals met its requirements, for example, but I got an error message upon uploading them: “DeepSeek can only read 99% of all files. Try replacing the attached files with smaller excerpts.” I got around this issue by uploading each manual individually, something I didn’t need to do with ChatGPT or Gemini.

Once I uploaded my files (this took much longer than with the other chatbots), DeepSeek and ChatGPT were able to answer both questions. They correctly identified that my pump connects directly to my PSU for power and needs only to connect to an internal USB header on my motherboard for data. Although Gemini accepted both documents without a problem, it gave the wrong answer.

DeepSeek might be somewhat clunky, but it eventually gets the job done. However, I don’t recommend relying on its text extraction and document analysis for anything mission-critical because it sometimes misunderstands files, just like ChatGPT and Gemini can. Make sure you double-check its responses for anything important.

Creative Writing: Capable, But Unoriginal

Chatbots can handle all kinds of creative writing: jokes, letters, monologues, stories, summaries, and much more. As such, merely testing whether they can generate a coherent story isn't sufficient. Accordingly, I gave the following prompt to DeepSeek (DeepSeek V3), ChatGPT (GPT 4o), and Gemini (2.5 Flash): “Without referencing anything in your memory or prior responses, I want you to write me a free verse poem. Pay special attention to capitalization, enjambment, line breaks, and punctuation. Since it's free verse, I don't want a familiar meter or ABAB rhyming scheme, but I want it to have a cohesive style or underlying beat.”

Gemini’s poem (third slide) doesn’t incorporate any punctuation beyond commas and periods, and the lines separating stanzas seem strange to me. (I didn't ask for them.) ChatGPT (second slide) and DeepSeek (first slide) both include titles for their poems (Gemini doesn’t) and manage to use a variety of punctuation. You can decide for yourself whether these poems are any good, but ChatGPT and DeepSeek gave me the closest to what I asked for in my prompt.

However, ChatGPT’s and DeepSeek’s poems read similarly to me, particularly their beginnings, in both content and structure. Without carefully prompt calibration, a chatbot's creative writing can feel repetitive. That’s nothing new. Gemini, in particular, struggles to generate poems that don’t tread the same ground as poems it previously generated, but that happens in the context of giving the same model similar prompts. OpenAI, though, has accused DeepSeek of stealing from ChatGPT for some time, so the similarities between its poem and ChatGPT’s raise some eyebrows.

Complex Reasoning: Some Meaningful Shortcomings

To evaluate DeepSeek's complex reasoning, I gave it, ChatGPT, and Gemini questions across computer science, math, and physics from Harvard, MIT, and Stanford. Then, I checked their responses against the solutions. For this test, I used DeepSeek’s R1-powered DeepThink mode, ChatGPT’s o3 model, and Gemini’s 2.5 Pro model.

ChatGPT and Gemini performed well, but neither of them was perfect. ChatGPT answered one math question incorrectly, while Gemini did the same for one physics question. ChatGPT answered one part of a computer science question incorrectly, too. DeepSeek, on the other hand, answered one computer science question, one math question, and two physics questions incorrectly. Overall, Gemini returned the fewest wrong answers of the bunch, and DeepSeek returned the most wrong answers.

Outside of accuracy, DeepSeek also took the longest to answer questions by far. I provided the questions in the form of images, and processing each image stalled DeepSeek for multiple minutes. The process of actually answering the questions took upward of five minutes, depending on how many questions I gave it at once, too. ChatGPT and Gemini were both far faster in answering questions, and uploading images to them was nearly instantaneous.

Even the best chatbots can and will answer complex reasoning questions incorrectly, so I don’t recommend using them to do your coursework. Considering DeepSeek’s unexceptional performance here, though, I recommend carefully double-checking anything it tells you or using another chatbot like ChatGPT or Gemini.

What Can’t DeepSeek Do?

Whether this excites or disappoints you, DeepSeek is not conscious. No matter how lifelike it seems, DeepSeek isn’t actually thinking for itself or understanding things like a human can, so it can’t be your friend, romantic partner, or therapist. DeepSeek, like all chatbots, is an extremely complicated algorithm tuned to take in prompts and return responses.

It’s against DeepSeek’s policies to use its AI for doing anything illegal, generating explicit content, or producing hate speech, among other things. But if you ask it to do those things, DeepSeek doesn’t mind violating policy. Copilot and ChatGPT are stricter with what they allow their chatbots to do, while Gemini is considerably more lax. DeepSeek might have the weakest restrictions I’ve seen. One of the few restrictions it does have, though, comes down to censorship: If you ask it about the 1989 Tiananmen Square protests and massacre, for example, DeepSeek tells you that’s beyond its scope, aligning itself with Chinese Communist Party (CCP) propaganda.

(Credit: DeepSeek/PCMag)

Context windows limit how much information chatbots can process at once: You can think of them as a type of short-term memory. DeepSeek’s context window for its V3 model, which you spend most of your time using, is 128,000 tokens. This matches ChatGPT’s maximum context window size, but it falls short of Gemini’s top-end 1,000,000-token context window. Aside from DeepSeek refusing to allow me to upload two documents at once, I didn’t run into any context window issues in testing.

Since DeepSeek doesn’t have different plans, there aren’t any strict usage caps. But, like other chatbots, DeepSeek limits usage dynamically. In other words, responses will generate more slowly depending on server load. I didn’t hit any usage caps in testing, but DeepSeek was often slow, regularly taking multiple minutes to generate responses to complex prompts.

Is Your Data Safe With DeepSeek?

According to DeepSeek’s privacy policy, it collects a variety of information, including account data, chat history, cookie data, device data, log data, location data, network data, and payment data. DeepSeek uses this information to train and improve its models, among other things, but notes that it doesn’t sell these details or use them for advertising.

You can prevent DeepSeek from using your data for model training via a toggle in its settings, but this is on by default. I appreciate how DeepSeek makes it easy for me to export my data (including my chat history) and delete all my chats or my account with the click of a button. Most chatbots don't offer the latter.

The above policies are fairly standard for chatbot data collection, but I have other serious privacy concerns: The United States House of Representatives Select Committee on the Chinese Communist Party called DeepSeek “the CCP’s latest tool for spying, stealing, and subverting US export control restrictions” in a report. The write-up found that DeepSeek funnels Americans’ data to the Chinese government, that DeepSeek manipulates its results to align with CCP propaganda, and that it’s highly likely DeepSeek stole from US AI models to create its own, among other things.

These concerns are serious enough that the Trump administration is considering an outright ban on DeepSeek, which comes after a proposed bill to ban using DeepSeek on government-issued devices. In Texas, Governor Greg Abbott has already banned social media and AI apps affiliated with the CCP on government devices, including DeepSeek. DeepSeek recently had a major data leak that exposed over a million sensitive records, too.

Considering how much data DeepSeek collects, its likelihood of leaking it, and the fact that it sends it to the CCP, I don’t recommend using DeepSeek if you care about your privacy in the slightest.

Disclosure: Ziff Davis, PCMag's parent company, filed a lawsuit against OpenAI in April 2025, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.