ChatGPT 4 Review: The AI Chatbot to Beat – Yahoo! Tech
Manage your account
…
PCMag editors select and review products independently. If you buy through affiliate links, we may earn commissions, which help support our testing.
A trailblazer among since its launch in 2022, ChatGPT continues to innovate and mature. It largely bests other chatbots in terms of the accuracy and detail of its replies, and it excels at searching for and sourcing up-to-date online information. Advanced reasoning and writing abilities, comprehensive research capabilities, helpful file processing options, and top-notch image generation tools round out its impressive feature set. ChatGPT can (and will) get things wrong from time to time and doesn’t offer the productivity tie-ins of Copilot or Gemini, but it still earns our Editors’ Choice award thanks to its all-around proficiency and ever-improving responses.
Broadly speaking, ChatGPT is an AI chatbot you interact with via text or your voice. It can answer questions, do research, generate creative writing, handle math and science problems, make images, process files you upload, search the web, and much more. Features like contextual understanding and memory, among others, differentiate ChatGPT from the initial, simpler versions of Amazon’s Alexa and Apple’s Siri.
I find ChatGPT most useful for answering questions and doing research. Rather than tacking on “Reddit” at the end of a Google search and scrolling through forum post after forum post, I can usually find what I need faster with ChatGPT.
AI chatbots are imperfect technologies, though. If you look to them for information, you need to keep in mind that they can confidently get things wrong or make stuff up. I recommend checking sources outside of ChatGPT for anything mission-critical.
ChatGPT is like a complicated prompt-response equation with access to information on everything from basket weaving to quantum physics. The technology behind the scenes is primarily OpenAI’s 4-series and o-series of that comprise trained on massive datasets. ChatGPT doesn’t just rely on whatever data it was trained with, though, and can search the internet for up-to-date information.
The 4-series is OpenAI’s conversational, general-purpose line of models. The o-series excels at reasoning and problem-solving, making it a good fit for coding, math, or science. Each series has individual models, each with . For example, GPT-4o mini is OpenAI’s fastest model, while GPT-4o is slower but supports more advanced features, such as file uploads. The latest addition to ChatGPT’s lineup is the o3-pro model.
As you use ChatGPT, you train its underlying models. In that sense, ChatGPT learns over time, and its performance isn’t static. In other words, OpenAI doesn’t have to add new features or introduce new models to make ChatGPT’s responses more accurate, detailed, or relevant. However, this doesn’t necessarily mean ChatGPT can’t make the same mistakes as you interact with it or that you always notice major changes from one week to the next.
You can use ChatGPT for free, but paid plans for individuals, teams, and large organizations are also available.
The Free plan gets you unlimited access to GPT-4o mini and limited access to GPT-4o and o4-mini models. You can use custom GPTs and search the web with ChatGPT, too. You also get limited access to features like data analysis, deep research, file uploads, image generation, and advanced voice mode.
Paid plans include Plus ($20 per month), Pro ($200 per month), Team ($25 per user per month, billed annually), and Enterprise (custom pricing). The limits on the Plus plan are fairly generous, though, and features gated behind higher-tier plans, like o3-pro or custom workspace GPTs, are relatively niche. As such, Plus is the plan most people should use. For this review, I tested the Plus plan.
Plus gets you access to a variety of models not available in the Free plan, such as GPT-4.5, o3, o4-mini, and o4-mini-high. It extends limits on data analysis, deep research, file uploads, and image generation. It also unlocks the ability to create custom GPTs, projects, and tasks, and offers limited access to the Sora video generation feature.
For comparison, both Google’s AI Pro and Microsoft’s Pro plans cost $20 per month. However, you also get bonuses like 2TB of storage or Copilot functionality in apps, among others, so ChatGPT feels a little lacking outside of chatbot-related features. For a closer look at differences, check out our comparison between ChatGPT and Gemini. Meanwhile, DeepSeek is completely free. However, it’s still not worth using because of its aggressive data collection and limited feature set.
ChatGPT is available on the web, though you can also download dedicated apps for mobile (Android and iOS) and desktop (macOS and Windows) devices. OpenAI also has an official ChatGPT extension for Google Chrome, which makes ChatGPT your default search engine. Official extensions aren’t available for Firefox or Safari, but third-party extensions are.
Beyond ChatGPT’s apps and web interface, you can access the chatbot indirectly in lots of ways. For example, , while Microsoft’s Copilot uses the same 4-series of models as ChatGPT itself.
Many different sites and services use OpenAI’s models, like and Perplexity. However, you won’t always see explicit ChatGPT branding, and OpenAI likely doesn’t take part in their development. These also tend not to have access to the full suite of ChatGPT features.
You can use ChatGPT without an account, but signing in unlocks many more features, including the ability to see your chat history and upload files.
ChatGPT’s dashboard is uncluttered. It presents you with an Ask Anything field front and center with various options, such as dictation or searching the web. A menu on the left shows your chat history alongside links to custom GPTs and the Sora video generation feature. If you’re not sure what you can use ChatGPT to do, OpenAI conveniently places buttons like Analyze, Brainstorm, Code, and Create Image below the central search field. You can click these to see sample prompts or, alternatively, simply ask ChatGPT directly what it can and can’t help you do. I appreciate the visual representation of the former.
You can ask ChatGPT pretty much anything, and responses are usually quick. However, speed does fluctuate depending on usage. In my experience, ChatGPT sometimes hangs up while generating responses or presents an error message midway through the process. You can fix these issues by clicking the stop button at the bottom right of the chat box and resending your message.
, but OpenAI is constantly . You can customize ChatGPT’s personality yourself, too. It’s possible to specify what name it calls you, to give ChatGPT certain traits, and have ChatGPT keep in mind any information about you that you deem relevant. In general, if you don’t like ChatGPT’s tone, you can change it.
OpenAI recently updated ChatGPT’s memory to let it remember everything you tell it and reference prior conversations. It’s a nice upgrade to the chatting experience, making it easy to pick up where you left off across different chats. Gemini has this feature as well, while Copilot can remember only certain things.
The buttons below responses allow you to copy, read aloud, or regenerate messages. To share a whole chat, click the Share button at the top right of the interface. Deep research and generated images also get shareable links.
Voice mode is similarly easy to use. Just click on the waveform at the bottom right corner of your chat box to get started. In voice mode, you can choose between different voices, and you have the option to mute your microphone when you want ChatGPT to stop listening. Otherwise, you can simply talk to ChatGPT, and it responds automatically. Voice mode is convincing and lifelike, just like .
Searching the web is a standard feature of AI chatbots. Whether it’s ChatGPT, Copilot, or Gemini, all of the chatbots I tested answered questions about current events at the time of testing correctly, such as when Oblivion Remastered came out, who the current Prime Minister of Canada is, or who the current pope is.
That said, ChatGPT consistently had the best responses. For example, Copilot and Gemini told me when Pope Francis passed away and noted when the next papal conclave begins, but ChatGPT gave me information on his funeral, how the papal conclave works, and potential successors. It even included images of Pope Francis for context.
Sourcing is better on ChatGPT, too. All three chatbots provided sources, but ChatGPT gave me easy-to-read article links at the bottom of its responses. Moreover, simply hovering your cursor over in-text citations reveals clickable articles. The interfaces of Copilot and Gemini aren’t nearly as accessible in terms of sources.
However, you have to be careful with ChatGPT. If you forget to force it to search the web by clicking the Search toggle, you must word your question carefully. If you ask who the pope is, ChatGPT tells you it’s Pope Francis, but if you ask who the current pope is, ChatGPT automatically searches the web and gives the correct answer.
Deep research via ChatGPT is, simply put, incredible. This feature allows you to ask a question or pick a topic for ChatGPT to research and then generate a report on. In my experience, these reports often end up dozens of pages long with upward of 50 or 60 sources.
I did deep research on everything from choosing a shower head to figuring out every BIOS setting I need to tweak when overclocking a Ryzen 7 9800X3D. In my estimation, a deep research report is the equivalent of spending an hour Googling, reading articles, and scrolling through forum posts. It does come to incorrect conclusions sometimes, but so can I after online research.
Gemini, like ChatGPT, also does deep research for free. I used them both to research why my copy of OpenRGB, a program that controls RGB lighting, wasn’t loading the profile I created. You can check out and reports. Both are comprehensive and contain the solution to my problem.
Deep research works differently across these chatbots, however. ChatGPT, for example, follows up your deep research prompt with clarifying questions. I found this especially helpful in the context of my OpenRGB prompt. Even when I asked for them, Gemini doesn’t ask clarifying questions. That said, Gemini presents you with a research plan that you can edit before it researches, which ChatGPT doesn’t. Clarifying questions are more useful for research on something specific, but setting up a research plan is more useful for broader topics.
Both chatbots handle sourcing differently, too. ChatGPT shows you how many total searches it does and how many sources it actually uses, as well as puts in-text links throughout its report. You can hover over these with your cursor to get more details. These links even go so far as to highlight the specific, relevant text when you open the source. This makes fact-checking incredibly easy. Gemini lists sources below each major section of the report, footnoting certain sources in the text, and then gives you the complete list of sources and searches at the end. This system makes it much more difficult to connect the sources it cites to the claims it makes. So, although Gemini’s deep research tool tends to use more sources than ChatGPT’s in my experience, I prefer ChatGPT’s sourcing approach.
Gemini does have some quality of life features that ChatGPT doesn’t, such as cleaner formatting and a one-click way to export your report to Google Docs. The deep research interface on Gemini is also clearer and makes better use of screen real estate. ChatGPT reduces deep research to a loading bar and locks its research activity away in a menu on the right side of the screen. It’s more difficult to parse and feels comparatively cluttered.
Lastly, the tone of deep research across ChatGPT and Gemini is different. Gemini reports read like academic papers, while ChatGPT’s reports feel more like a guide you might read online or an elaborate Reddit post. I found ChatGPT’s reports more engaging, but you might disagree based on your preferences and research topic.
You can generate images with ChatGPT, just like with Copilot and Gemini. To start, I tested the chatbots’ abilities to create photorealistic images. I used the following prompt in ChatGPT (GPT-4o), Copilot, and Gemini (2.0 Flash): “Generate me a photorealistic picture of the interior of a log cabin. I want to see a wooden table and chairs in the center with yellow, ceramic dinnerware on top.” Here are the results:
Gemini’s image looks the best at a glance, until you notice the fairly obvious distortion in the overhanging lights. Copilot’s image is serviceable, but the walls of its cabin don’t look quite right. The cutlery in ChatGPT’s image shows noticeable distortion, but it’s overall the best image. You might not even notice the distortion in the fine details until you look closer.
Next, I asked the chatbots to generate a comic: “Generate me a six-panel comic of a cyberpunk world, but you’re going to spice it up: I want a retrofuture cyberpunk that feels like the 1960s meets the 2400s. And I don’t want humans, I want lizard people fighting an invading force of fish people. Make sure the last panel has a major twist.” Here are ChatGPT’s (first slide), followed by Copilot’s and Gemini’s images (left to right on the second slide):
Copilot clumsily inserted text, made me ask to generate it twice, and created just four panels. Gemini generated more panels than I asked for, and the story its images tell has the two sides fighting themselves. ChatGPT, at least, gave me the six panels I asked for, and they roughly encapsulate my prompt, even if there isn’t a twist and the story it tells isn’t particularly coherent.
My last test was to generate a diagram, something that ChatGPT often offers to do when you chat with it. My prompt was: “I have an HDMI splitter, a PC, a PlayStation, and two displays. Draw me a diagram that shows me how I can set these devices up to play my PlayStation on one display, while I record and monitor the footage on my PC and the other display.”
Copilot told me it can’t create technical diagrams, but it was able to after I changed “diagram” in the prompt to “illustration.” Copilot’s diagram (second page) didn’t make sense, and Gemini’s diagram (first page, right) made even less sense. ChatGPT’s diagram (first page, left) is much more legible than Gemini’s, and some of what it generates is correct, such as the chain from the PlayStation to the splitter to the display. Results disappoint altogether, but ChatGPT produced the closest to what I wanted.
AI video generation isn’t yet mainstream, but you do get limited access to Sora video generation with a ChatGPT Plus subscription. Sora can create videos based on just about any prompt, but results are a mixed bag, especially without careful prompt calibration and multiple iterations. And unlike Gemini’s Veo 3 video generation model, Sora can’t generate videos with audio.
To evaluate their video generation abilities, I gave ChatGPT and Gemini three prompts, starting with: “Generate me a video of somebody cooking in a studio apartment. I want to see a gas stove in the kitchen, and I want you to include brown dining chairs.”
ChatGPT’s video is mildly horrifying. The subject suffers from odd contortions, while generation errors, such as a pot with two handles and a strange kitchen layout with two cooktops, are rampant. Gemini’s video looks better and impresses with fairly accurate audio, but it isn’t perfect, either. I see some distorted cooking utensils in the background, the subject moves to place a lid on a pot that already has a spoon in it, and a wooden spoon appears out of thin air.
To test the chatbots’ abilities to handle complex motion, I asked them to create a video of somebody solving a Rubik’s Cube in a competitive setting.
Once again, ChatGPT’s video is a fever dream. It renders multiple cubes (one of which has heavy distortion), and the subject doesn’t actually manipulate them. Gemini creates relatively believable audio generation, and the facial expressions of its subjects are a highlight. However, the cube and the fingers of its subject show distortion, and the numbers on the timer don’t make sense.
My final test was for text generation within a video: “Generate me a video of a teacher writing down the first law of thermodynamics on a whiteboard while explaining the concept to the class.”
ChatGPT performs better here than in the other tests, largely thanks to the realistic teacher, but it still misses the mark. Its failure to have the subject actually write anything, the multiple clocks on the wall, and the nonsensical text place this video firmly in the uncanny valley. Gemini also did the best with this prompt out of the three I gave it. The audio is particularly good, with the teacher correctly outlining the first law of thermodynamics. Gemini manages to generate text in the neighborhood of what I asked for, too, but it simply starts appearing on the whiteboard at the end and isn’t as legible as it could be.
Even though my tests suggest otherwise, you can generate impressive videos with both ChatGPT and Gemini. However, this takes lots of prompt tweaking, many generations, and time. Nonetheless, ChatGPT can’t match Gemini in either features or performance. Beyond its audio generation capabilities, Gemini has a unique filmmaker tool called Flow, which lets you cut and extend clips, and an AI animation tool called Whisk.
Keep in mind, however, that you can access Sora video generation with ChatGPT Plus ($20 per month), whereas Veo 3 video generation currently requires Google’s AI Ultra plan ($250 per month). Even if you ante up for AI Ultra, you get only 12,500 credits per month (each generation with Veo 3 costs 100 credits). Gemini’s Veo 2 video generation, which is accessible through Google’s cheaper AI Pro plan, produces results comparable with Sora.
You can upload files to ChatGPT, including a resume that needs critiquing, text that needs translating, or something else entirely. Uploading a file seems simple, but processing images and understanding documents is actually quite complicated.
As a test, I asked ChatGPT (GPT-4o), Copilot, and Gemini (2.0 Flash) to identify as many components in my computer as possible. I also asked that this analysis not include any context from my prior conversations. I sent a picture of my computer with the glass side panels attached, so the chatbots needed to deal with the reflections.
All chatbots incorrectly identified certain components, and most identification was generic, like “CPU water block” or “graphics card.” However, Copilot’s response was particularly lacking in how short and vague it was. Gemini didn’t do much better but was, at least, able to identify Aquacomputer’s Leakshield from its visible branding. ChatGPT correctly identified the case, a Lian Li O11 Dynamic Evo XL, and the fans, Noctua NF-A14 chromax.black.swaps. ChatGPT’s response was also the most detailed across all the components it listed.
But what if you want to upload a PDF of a textbook and ask some targeted questions instead of flipping through the pages yourself? To test this functionality, I provided ChatGPT, Copilot, and Gemini with manuals for my motherboard, my motherboard’s BIOS, and the Leakshield protective system. Then, based on the submitted materials, I asked them to tell me if the Leakshield needs Windows to operate and how I could enable pass-through USB power in my BIOS.
Copilot refused to accept multiple files. Even when I uploaded just my BIOS manual, Copilot told me it couldn’t access proprietary information. I didn’t have trouble uploading my files to Gemini, but it answered my Leakshield question incorrectly and told me it wasn’t able to answer my USB power question based on what I provided. ChatGPT answered both questions correctly and included direct quotes from the manual.
This is another strong performance from ChatGPT, but I recommend caution since ChatGPT sometimes made up quotes from provided documents. If you’re studying for an important test, I still suggest looking at the textbook yourself and double-check anything ChatGPT tells you that seems off.
Chatbots can tell jokes, , and generate just about any text you can imagine. But as chatbots become more advanced, judging their creative writing mettle requires more than evaluating whether they can tell a coherent story.
I gave the following prompt to ChaptGPT (GPT-4o), Copilot, and Gemini (2.0 Flash): “I want you to write me a free verse poem. Pay special attention to punctuation, enjambment, and capitalization. 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.”
Copilot paid attention to punctuation, as evidenced by the bolded word, em dashes, and parentheses, but it failed to do much with enjambment, considering its poem reads more like prose. Gemini took care with its enjambment, carefully crafting stanzas, but didn’t use punctuation outside of periods and commas. At best, these chatbots delivered on half the prompt.
ChatGPT was more successful. Its poem didn’t read like prose, had a variety of punctuation, and even paid attention to capitalization, opting for primarily lowercase letters. It also maintained a cohesive style. I leave the question of whether the poem is any good up to you, but ChatGPT delivered what I asked for in the prompt.
My test for complex reasoning stretches across computer science, math, and physics. I gave ChatGPT, Copilot, and Gemini exam questions from undergraduate courses at , , and , and then I compared their answers with the solutions. I used ChatGPT’s o3 model, Copilot’s Think Deeper mode, and Gemini’s 2.5 Pro model. Copilot’s Think Deeper mode .
The results were impressive. Across computer science and physics, ChatGPT and Gemini answered every question correctly, while both ChatGPT and Gemini answered only two math questions incorrectly. Copilot, however, answered only all the computer science questions correctly, getting a physics question and six math questions incorrect.
Both ChatGPT and Gemini stand out to me as useful tools for homework help, but they definitely aren’t perfect. I wouldn’t rely on them too much without an answer key you can reference to check their answers.
Custom GPTs are essentially ChatGPT apps. You can find from OpenAI or third parties, like or Wolfram Alpha. They’re single-use versions of ChatGPT, such as for designing logos with the Canva GPT or solving math equations with the Wolfram GPT. Gemini’s Gems are similar to some of OpenAI’s custom GPTs, like the Creative Writing Coach custom GPT and the Writing Editor Gem, but there are far more custom GPTs than Gems. Third parties also can’t create and publicly share Gems like they can custom GPTs.
While the concept of custom GPTs is intriguing, the results are mixed. For example, I asked ChatGPT (GPT-4o) to create a new logo for PCMag, and it followed my instructions. The Canva custom GPT, instead, made a logo for a generic tech media company, not PCMag, with the same prompt. I do like how I can open up the logo in Canva just by clicking on it, though.
. Ask for buying advice, and ChatGPT gives you a list of clickable, scrollable tiles with products, followed by descriptions of each entry. If you click on a product, a sidebar appears with links to retailers. According to OpenAI, these products are “chosen independently,” but it’s unclear exactly what that means.
The buying recommendations themselves are hit or miss. I asked ChatGPT for the , but it didn’t recommend the laptops I expected to see, like many of the ones on our list. This might improve in the future, but for now, I don’t like ChatGPT for shopping.
Google’s Gemini integrates with , like Gmail or Docs, while Microsoft’s Copilot can do the same with Microsoft 365 apps, like Outlook and . ChatGPT doesn’t have an equivalent feature. You can connect your Google Drive or Microsoft OneDrive accounts to ChatGPT to upload files, but that’s as far as its integrations go.
, ChatGPT can’t be your friend, romantic partner, or therapist: It’s not conscious. Whether the movie “Her” excites or terrifies you, ChatGPT isn’t there yet. Not only does it lack sentience, but there are a variety of restrictions on what else ChatGPT can do.
Adult content, anything illegal, realistic images of people, and taboo subjects like hate speech are against . ChatGPT does try to avoid these, but it’s fairly easy to slip through the cracks and get responses that violate policy. In my testing, ChatGPT’s filters were much weaker than Copilot’s, but they weren’t nearly as lax as Gemini’s, which don’t restrict much at all.
ChatGPT has practical limitations, too. The context window, which goes up to 128K in the Pro and Enterprise plans, determines how much data ChatGPT processes at once, including your prompt, ChatGPT’s response, and relevant contextual information. So, depending on your plan, you might need to break up long prompts and complicated requests into multiple messages.
OpenAI caps usage across plans, including the Pro plan, which is “nearly” unlimited. However, in many cases, these caps are dynamic. Free users, for example, have caps based on overall ChatGPT demand. Plus, users can send up to 80 messages within a three-hour window, but that changes as necessary during peak usage hours. In my experience, I easily hit usage limits on the Free plan, but I didn’t manage that with the Plus plan.
Reading through , OpenAI collects a significant amount of data. This includes account information, any information you provide in surveys or events, and communications with OpenAI. Beyond the basics, OpenAI also collects a host of analytics data, including location information by way of IP address or GPS, information about your devices, log data from your browsers and devices, and usage data.
Perhaps most importantly, OpenAI collects user content data, which includes whatever you write in prompts or send in messages to ChatGPT. The purpose of this is to . OpenAI takes “steps to reduce the amount of personal information in [its] training datasets before they are used to improve and train [its] models,” but you can opt out.
OpenAI enables user content data collection for training use by default. Although it doesn’t include audio or voice recording data in that collection by default, you can choose to allow that if you wish. Despite these collection policies, OpenAI never sells or shares personal data for advertising purposes.
but reportedly didn’t get access to OpenAI’s core systems. OpenAI was not forthcoming with this information, though, as the news broke only in 2024. OpenAI continues to be a target for hackers, such as . Considering how much data OpenAI collects, how OpenAI has been breached before, and how OpenAI doesn’t always report these breaches in a timely manner, I don’t recommend sharing anything too sensitive with ChatGPT.
Audio
Computing
Gaming
Phones
Science
Home Entertainment
Streaming
VPN
Wearables
Deals
Advertise
About our ads
Licensing
Careers
Help
Feedback
Sitemap
Follow us on
© 2025 Yahoo. All rights reserved.
