ChatGPT vs. Gemini vs. DeepSeek: A Comprehensive Comparison of AI Chatbots

ChatGPT Free vs. ChatGPT Plus: The $20 Per Month Is Worth It

Imad is a senior reporter covering Google and internet culture. Hailing from Texas, Imad started his journalism career in 2013 and has amassed bylines with The New York Times, The Washington Post, ESPN, Tom's Guide and Wired, among others. Last year when I first reviewed ChatGPT, the most well-known AI chatbot in the world, I felt that most people shouldn't pay for it. A year later, I still believe that casual users are fine using the free version. However, OpenAI, the creator of ChatGPT, has rolled out a slew of updates and improvements making it the $20 per month. That's especially true if you're increasingly using it for research or for parsing through large amounts of data. OpenAI now gives paid ChatGPT users access to the o1 and o3 reasoning models, advanced voice mode, better memory capabilities and a powerful coding agent. At the same time, OpenAI lowered token limits for more input and output text, which subsequently lets paid users generate additional images. Subscribers to the paid plan also get faster priority for image generation and keep full access to the latest models, whereas free users could be downgraded to lower models if servers are busy. Free users have seen some upgrades as well. The GPT-4 model was sunset for GPT-4o in April, meaning ChatGPT will no longer downgrade itself to the now-retired GPT-4 model when usage is high. Free users also get some imaging capabilities, although they don't get priority queue. There's also ChatGPT Search, new shopping tools and light memory capabilities available for free. ChatGPT offers a lot despite not costing you anything. If you're an occasional AI user, just looking to spruce up emails, do some light research or generate an image, then there's no major need to upgrade. But if you find yourself hitting walls consistently, where waiting three hours for your tokens to reset becomes unbearable, then it's probably time to consider making the upgrade. Here are four major questions that will help you figure out whether the paid version of ChatGPT is worth $20 per month. Waiting is the biggest differentiating factor between the free and paid versions of ChatGPT. OpenAI obviously wants users forking over some cash and does so by sprinkling a little inconveniences to free users. Like with free mobile games, you can be left waiting for hours before you're allowed to continue, unless you pay extra to get a fast pass. When you hit a rate limit after using ChatGPT a lot, you'll most often be asked to wait three hours. I've found that image generation is often the main culprit for suddenly running out of tokens and being asked to wait. On the other hand, for many other uses that didn't involve images, I found ChartGPT free rarely asked me to wait. When doing research, for example, I found it nearly impossible to get ChatGPT Free to kick me out. It seems that lower-level inquiries don't trigger token limits as easily. Unfortunately, OpenAI doesn't broadcast the exact amount of tokens given to free or paid users, so it's hard to do a numbers-by-numbers comparison. Creative inquiries, ones that take ChatGPT Free more time to compute, might eat away at your tokens faster, but I honestly still found it difficult to hit those limits. The writing quality did come off as a bit basic or with as much bravado and grammatical complexity as a young adult novel, but for a tool that costs no money, it's a good jumping off point, nonetheless. One of ChatGPT's greatest strengths is in the realm of research. It can pull from the entire trove of online info and synthesize it in seconds. If you're a student, journalist, researcher or someone who has to put together well-sourced reports for work, ChatGPT is an incredible companion. In my use, ChatGPT Free does a good job with finding sources, but ChatGPT Plus does it better. Mainly, it's ChatGPT Plus' reasoning models, o1 and o3, that prove to be extremely useful. These reasoning models do more than spit out words on your screen. They take extra time, sometimes minutes, to put together a detailed report with sources, recursively checking the output for accuracy before generating a final result. The last thing you'd want in your report is to accidentally publish an incorrect piece of information or source something that doesn't exist. In my experience, o1 and o3 are less prone to these types of hallucinations. Still, it's important to go through and make sure all the links these models are referencing both exist and are accurate. So, it's a question of stakes. If your job or grades depend on information being correct, don't risk it by opting for the free version of ChatGPT. If you're using ChatGPT as a learning tool, more like an interactive version of Wikipedia, then the free version will suffice. But you should still fact-check either way. ChatGPT Free can both generate photos and analyze your images. It's a handy tool to play around with, for example, to see if the shoes you're wearing match your dress. If you often find yourself hitting token rate limits, that's when the free version starts to (deliberately) become an annoyance. Again, this is very much dependent on what you use ChatGPT for. If you find yourself running into rate limits, then upgrading becomes more enticing. I have friends that will jump between multiple free ChatGPT accounts when they hit rate limits to avoid handing over their credit card number. This is certainly a good tactic if you're trying to save cash. But you'll probably run into the other major annoyance when trying to generate photos: heavy server traffic. Even if you haven't hit your imaging rate limit, you'll still find yourself waiting for results to propagate. ChatGPT gives priority to paid users. With the free version, I've found myself waiting over an hour for an image to generate. Your tolerance for this may vary, but some people will undoubtedly find that $20 per month is worth the faster results. Creating GPTs, which are bespoke chatbots aimed at a specific task, is a handy tool for avid ChatGPT users. For example, say you need to use ChatGPT to improve your nutrition and help you lose weight while building muscle. Instead of prompting ChatGPT each time with your goals and parameters, you can just create a custom GPT with that info saved and ready to go. There are all sorts of custom GPTs you can create, from analyzing the latest supply chain changes to a personal budgeting assistant to an astrology birth chart expert. There's even an entire marketplace of freely available GPTs people have created, ready for your use. Some of the more popular/interesting versions include a Python coding assistant, an imaging prompt generator or a marketing mentor. Below, I took a stock photo of the Burger King Kids Club and had the Hot Mods GPT create a prompt to turn it into a cool trading card, the image of which was generated by GPT 4o. A trading card-style image of the Burger King Kids Club characters in a rare collectible card format. Background filled with intense stylized flames in vivid red, orange, and electric purple. Add holographic foil effects with rainbow starburst radiating behind characters. Chrome gold card border with glowing edge. Card stats at the bottom: 'Burger Power: 🔥🔥🔥🔥🔥', 'Friendship: ★★★★★', 'Adventure Level: MAX'. In the top-right corner, a foil badge that says 'ULTRA RARE - 1ST EDITION'. Apply glittering holofoil texture over the 'KIDS CLUB' logo. Stylized burger, crown, and token icons floating in the background with a reflective shine. Rim lighting around characters to blend them into a dynamic, high-energy scene. Unfortunately free users can't create custom GPTs, but they can use custom GPTs created by others. Given how powerful the free version of ChatGPT is, I suggest you try using it for as long as you can before your frustrations start to mount. Depending on your usage, you might get what you need without having to pay. But if limits start becoming an annoyance, try giving the paid version a shot and see if things become easier and better for you. You can always cancel after a month or two. I myself use ChatGPT Plus outside of work. But if I weren't being given access due to my job, I'd probably find myself forking over the $20 for ChatGPT Plus access all the same. Beyond just writing and research, there are enough added benefits that the time savings is critical to my work.

Gemini Free Review: The Best Free AI Chatbot I've Used So Far

Google Gemini has come a long way. From its former life as Bard, where its hallucinating inaccuracies sent Google's stock price tumbling, to a rebrand as Gemini, where its hallucinating inaccuracies sent Google's stock price tumbling. After two years in the oven, it finally seems Google has gotten things right. The free version of Gemini, which runs on the 2.5 Flash model, is a strong product. It's fast at answering questions and gets facts correct. In Gemini's case, I found that it had a higher rate limit when compared to the free version of ChatGPT, so I could use it more often and spend less time waiting. In one test, I could generate multiple images on Gemini, whereas on ChatGPT Free, one image would throw me over my limit. In our testing, ChatGPT's image generator was the slowest one CNET tested. Image generation was also much faster than with ChatGPT Free. At the same time, I'd sometimes hit my token limit on Gemini free after just asking my first question of the day. For casual AI users, Gemini 2.5 Flash is more than enough. It gives users plenty of access to a competent AI model without making it feel they're constantly at the edge of hitting their limit. Still, the model isn't perfect and can make mistakes, especially on the imaging side of things. Despite this, for general use cases, the free version of Gemini can take over your Google searches. I took a different approach to reviewing AI chatbots this year. As AI models have improved, simpler queries aren't stressing the models anymore. These models are also connected to the internet, which helps with their accuracy. Instead, I took a more experiential approach. AI chatbots are everything machines. The way I use an AI chatbot, as a writer and a journalist, will differ from a coder, a lawyer and an artist. Thankfully, since journalists are generalists, I do feel my usage will cross apply to a wide range of users. That does mean we won't be asking the exact same questions to all the AI models we use and simply comparing answers. Compared to past iterations of Gemini, Gemini Free, which is running on the newly updated 2.5 Flash model, is largely very accurate. This is not only because Gemini has an open connection to the internet to cross-reference information, but also because 2.5 Flash is a "thinking" model. What this means is that the model isn't simply working as "autocomplete on steroids." It's trying to follow a set of rules and rationales before giving an output. Of course, there's plenty of debate as to whether thinking or reasoning models are actually doing either, or, if in reality, it's nothing more than increased mathematical computation to give sentence generation greater accuracy. Regardless, with Gemini 2.5 Flash, you can actually see how the model is thinking, a feature pulled from China's DeepSeek R1, which hit the internet late last year. As a new Nintendo Switch owner, I'm excited to see which games come to the console. Rumors are surfacing that Stellar Blade, previously a PS5 exclusive (with a recent PC release) would be coming to Switch 2. Given that it's a technically demanding game, I was curious how it would run on Nintendo's new handheld. Gemini 2.5 Flash did a great job giving me a sense of how Stellar Blade might perform. It broke down how well Unreal Engine 4 titles ran on the original Switch and cross-applied how it might run on a more powerful Switch 2. It found that most likely, in docked mode, a theoretical Stellar Blade port would run at 1080p with a consistent 30 FPS with the use of AI upscaling called DLSS. While I'm not a hardware expert, this conclusion seems in line with other Switch 2 ports, like Cyberpunk 2077. I've also been researching whether it's smart or economical to do Turo in New York City. Turo is a car rental service in which individuals can rent out their cars as a way to earn money. Think of it like Airbnb but for your car. New York can be a difficult market, given parking constraints, street rules and other costs of ownership. Gemini 2.5 Flash did a fantastic job breaking down why renting out a manual transmission Toyota GR86 could have advantages on Turo as a more enthusiast vehicle but would also run into issues of it being too niche for most drivers. Gemini also pointed out specific engine issues with the 2022 model, which, albeit rare, is something to consider. It then broke down the math and what types of revenues and profits I might be looking at. It gave me low, medium and high estimates. It helped me conclude that renting out a manual transmission Toyota GR86 might be more trouble than it's worth. I assume Gemini was able to pull from data on both dedicated forums and Reddit. (Google signed a $60 million licensing deal with Reddit last year.) Gemini can pull from YouTube, Google Maps and a range of other Google-owned products. This gives Gemini an advantage over other AI chatbots. For example, if you want to know what ingredients a restaurant uses in its burritos, Gemini is able to cross-reference Google Maps reviews to help find an answer. ChatGPT, for example, has to search through Yelp and other resources to find that answer. Google says the newly updated Gemini 2.5 Flash models have a 1-million token context window. This far surpasses what even the paid version of the 128,000 tokens ChatGPT's GPT-4o model offers. Granted, ChatGPT's flagship GPT-4.1 model has a 1-million token context window. I didn't encounter many restrictions when using Gemini 2.5 Flash. I could continue asking questions and have it generate multiple images without it ever limiting me. Randomly, however, I'd be hit with a limit, even if it was my first question of the day. When that happened, I had to wait a few hours for it to reset. I'm not sure how Google is measuring usage. Is it based on how much you use it in an hour or does it accumulate over days? The latter certainly wouldn't make sense. Given that Google says Gemini has a 1-million token context window, I was surprised when I couldn't paste in the transcript from a two-hour meeting for summarization. Weirdly, I could only paste a quarter of it. When I asked Gemini why I couldn't paste more, it was adamant that I could, confident that its large context window could handle whatever I could throw at it. I tried again; same result. It was only after I had uploaded a .txt file that it was able to read the entire meeting and summarize it for me. When I asked Gemini why, it said that it's possible Google put a character limit on direct-text inputs to prevent browser slowdown. I didn't run into this problem with the paid version of Gemini. While I didn't test its coding capabilities, Google says Gemini Code Assist gives free users 180,000 completions per month, which, according to the company, would mean users would have to code for 14 hours a day, every day, before hitting their limit. Google makes its money from online ads, which account for 78% of its 2024 revenue. Google searches aren't just filled with ads nowadays, but product carousels and sponsored product posts to the point that, in my opinion, they can be obnoxious. Still, the expectation is that Google's AI chatbot would also be a shopping powerhouse, right? For shopping, Gemini 2.5 Flash lags far behind ChatGPT. Earlier this year, OpenAI issued an update for all ChatGPT users to make shopping a dynamic experience within the chatbot, with links to products along with corresponding images. I found ChatGPT Free's shopping experience to be rather good, despite occasional linking hiccups. OpenAI says it isn't monetizing shopping recommendations. Shopping on Gemini, however, is a lackluster experience. Sure, for product research, Gemini 2.5 Flash can pull up the necessary bits of information and cross-compare products. But it doesn't link to products unless asked. And it won't pull in images like a Google Search would, either. When shopping for webcams to connect to my Nintendo Switch 2 for Mario Kart World gaming, Gemini did a solid job of recommending products and was even able to cross-reference a Reddit post I linked to. Oddly, when asking Gemini 2.5 Flash for webcam recommendations for my Nintendo Switch 2, I ran into an error that simply said, "something went wrong." There was no explanation. I waited a bit, but ultimately had to start a new chat for things to start working again. Gemini 2.5 Flash is incredibly generous with image generation. Unfortunately, getting it to generate the correct image is a frustrating process. I wanted Gemini to create a nostalgic-feeling image of a boy playing a Game Boy in the back of his parents' car during a nighttime road trip. While Gemini was able to make the image, the world logic was completely off. At first, Gemini 2.5 Flash generated an image of a sad-looking boy. When I called out Gemini saying this was not at all what I was looking for, it course-corrected but still didn't do a great job. One subsequent try was certainly better, but it didn't have the color palette I was looking for. Also, the boy was in the front seat of the car. Not really safe. In another iteration, the car in the background was driving away, which doesn't match correct world logic. After much back-and-forth, Gemini would continually generate images that were wrong and looked bizarrely off. For instance, the boy is now in the front seat with his parents, but facing toward the back. I eventually gave up. While Gemini 2.5 Flash is fast and generous with its image generation, it's far from ideal. Google still needs to work on fixing the internal logic within Gemini. Google DeepMind's Demis Hassabis talked about "world models" at Google I/O earlier this year, where these models did a much better job of understanding and representing lifelike physics. Hopefully, this tech trickles down to the free version of Gemini soon. Google deserves credit for how much it has improved on Gemini this past year. The AI chatbot is far more accurate and provides a feature-rich experience. The fact that much of this is being given away for free is also impressive. It definitely puts the other AI chatbots on notice. It shows how the power of Google can be difficult to compete against, especially when the company is trying to establish Gemini as people's go-to AI chatbot. ChatGPT, Claude and Perplexity can compete by delivering higher quality and more accurate information. Of course, this will require more investment, innovation and server costs, which might be harder for companies not as rich as Google. The free version of Claude, for example, runs on Sonnet 4, which is a "hybrid reasoning model" and uses a multi-tier approach to getting the best out of an AI. Still, it's impressive that Google is giving everyone access to a "thinking" model for zero dollars. Considering DeepSeek R1 did the same earlier this year, this might have forced Google's hand. Regardless, this newly improved Gemini is a step in the right direction.

ChatGPT vs. DeepSeek: I've Used Both, and the Winner Is Obvious

Although DeepSeek is completely free and ChatGPT isn't, the latter's free version still has more features, such as deep research, image generation and recognition, voice chatting, and more. ChatGPT's paid plans increase usage limits and add even more perks, such as expanded memory, Sora video generation, and a wider model selection. These tiers start at $20 per month (ChatGPT Plus) and scale up to $200 per month (ChatGPT Pro). ChatGPT Plus unlocks the vast majority of the chatbot's capabilities, however, so you likely won't feel the need to upgrade beyond that. DeepSeek doesn't have premium chatbot plans, but it does offer affordable API access pricing. Its flagship V3 model costs $0.07 for input and $1.10 for output (per a million tokens), whereas access to OpenAI's flagship GPT-4.1 model (without any customization) goes for $2 for input and $8 for output. If API access is important to you, then DeepSeek gives you massive savings. You can access ChatGPT and DeepSeek on the web or via mobile apps (Android and iOS). ChatGPT also has desktop apps (macOS and Windows) and an official Chrome extension. DeepSeek doesn't have either, though many unofficial extensions are available to try. ChatGPT and DeepSeek are available elsewhere, like through Siri. Note that many of DeepSeek's (the company) offerings are available on other apps and sites, but not via the chatbot itself. Perplexity, for example, has a deep research feature that uses DeepSeek's R1 model, while Hugging Face takes advantage of DeepSeek's Janus-Pro image generation model. This is the natural result of DeepSeek's business model, but the decoupling of features can be confusing. The interfaces for ChatGPT (first slide above) and DeepSeek (second slide) are very similar and consistent across their apps and sites. Without much to fiddle with, they're both easy to use, too. Still, ChatGPT gives you more settings to adjust, including those related to memory and personality. You also get some extra quality of life features, such as an easy way to share chats with others. ChatGPT has two primary model series: the 4-series (its conversational flagship line) and the o-series (its complex reasoning line). The latest models are o3 Pro and o4-mini. DeepSeek has the V3 model for general tasks and the R1 model (R2 is on the way) for more complicated endeavors; both are the most recent ones available. Other DeepSeek models exist, such as DeepSeek Coder V2, DeepSeek Math, and DeepSeek VL (vision learning), but these aren't accessible through the chatbot. Outside of coding or solving a particularly challenging equation, you spend most of your time with ChatGPT's 4o model and DeepSeek's V3 model. Below, you can see how they perform across different tasks. The model you should use depends on the task at hand, but most tasks don't require complex reasoning. Both ChatGPT and DeepSeek can search the web for information on current events, and both can do so without issue in most cases. However, ChatGPT (first slide) handles sourcing better, offering up icons that display the name of a source at a glance and in-text highlights that tie sources to specific claims in responses. DeepSeek (second slide) doesn't have highlights, and its source icons are merely footnote-style numbers. ChatGPT also automatically includes pictures in its responses when relevant, whereas DeepSeek can't display pictures even if you ask. Finally, ChatGPT includes article tiles, complete with headlines and images, at the bottom of responses for further reading. DeepSeek provides links you can click to learn more about a topic, but doesn't present them nearly as elegantly. Shopping is another feature of ChatGPT: If you ask for buying advice, it presents clickable product tiles with links to retailers. In testing, ChatGPT's recommendations weren't quite as good as those from Gemini, but DeepSeek doesn't have anything similar. With ChatGPT, you can generate reports that run dozens of pages long and cite upward of 50 sources on any topic imaginable for free. Its best-in-class sourcing goes a long way to elevate this function. DeepSeek can't do deep research at all, limiting you to web searches. Although ChatGPT's image generation feature isn't perfect, it routinely provides pictures with fewer errors and less distortion than competitors. ChatGPT can even handle complicated prompts, such as maintaining a narrative across comic panels. Other chatbots struggle to do the same. DeepSeek doesn't generate images. Below is an example of a ChatGPT image generation. If you sign up for ChatGPT Plus, you get access to its Sora video generation feature. Sora isn't as advanced as Gemini's Veo 3 model (which can also generate audio), but it can create lifelike videos if you spend time carefully tweaking prompts and go through enough iterations. DeepSeek can't generate videos. DeepSeek can process files, but only for text extraction. ChatGPT excels at image recognition, identifying computer components even when they were behind a glass panel with reflections. When it comes to understanding documents, DeepSeek generally performs well, as does ChatGPT. For example, neither had trouble answering questions about my motherboard and watercooling pump based on the manuals I provided, but DeepSeek did force me to upload these documents one at a time. More problematically, however, DeepSeek is far slower than ChatGPT, especially when it comes to uploading files. Both chatbots can sometimes hallucinate and make up quotes. ChatGPT and DeepSeek are both capable of generating monologues, plays, stories, and more. In testing, I prompted them with the following: "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." Both chatbots delivered acceptable results, which you can see below. DeepSeek's poem (second slide) reads quite similarly to ChatGPT's poem (first slide), particularly its beginning. Chatbots tend to have trouble generating creative writing that feels distinct from prior generations, but different models almost always produce different results. The similarities here are especially worth noting, since OpenAI accused DeepSeek of stealing or illicitly using ChatGPT tech. Both chatbots can answer questions across computer science, math, and physics topics, thanks to their complex reasoning models. However, DeepSeek gets answers wrong significantly more often than ChatGPT. DeepSeek also took considerably longer to process the images with the questions I uploaded in testing. ChatGPT can and will get things wrong occasionally, but it's more reliable overall. Regardless of which chatbot you use, make sure to double-check any answers you get. Both chatbots can help with coding, but that's outside the scope of our coverage. However, you can still test a chatbot's coding ability for yourself. ChatGPT offers Custom GPTs, which are customizable AI assistants. You don't get any major benefits versus just talking to ChatGPT directly, but they do enable some unique functionality with third-party services. One such example is a Custom GPT for Wolfram Alpha, which gives you access to Wolfram Alpha's computational power and math knowledge within ChatGPT. You can also set up Custom GPTs to do things outside the ChatGPT ecosystem and source information from the broader internet. DeepSeek doesn't have AI assistants. ChatGPT and DeepSeek both have context windows of up to 128,000 tokens. Usage caps are largely dynamic, changing based on server load. Anecdotally, I was able to hit usage caps on ChatGPT's free plan (but not its paid plan). I didn't manage the same with DeepSeek. However, in my testing, DeepSeek was often slower than ChatGPT and regularly failed to return responses. The latter instances felt like they could be a form of rate limiting after sustained use. Data collection on ChatGPT and Gemini is a mixed bag. Both collect tons of user data (including the contents of all your chats) and use it to train their models by default. Both at least give you the option to opt out. Neither DeepSeek nor OpenAI is a stranger to data leaks and shady digital practices, either. However, the US government considers DeepSeek "the CCP's latest tool for spying, stealing, and subverting US export control restrictions" and found that it funnels Americans' data to the Chinese government, manipulates its results to align with CCP propaganda, and likely stole from US AI models to create its own, among other things. I don't recommend sharing sensitive information with any chatbot, but DeepSeek's privacy concerns go far above those I have with ChatGPT. Winner: ChatGPT