Real-world testing of ChatGPT (latest version 2026): Is ChatGPT still worth it?

It’s 2026, is ChatGPT still worth the hype?

In today’s world where large AI models are all the rage, OpenAI’s ChatGPT is nothing new, yet it’s still hailed by many as the “ceiling of AI.” But frankly, after using it for a while, you’ll find it exudes a coldness and arrogance—unlike domestic models that are considerate of Chinese users, and lacking the flexible adaptation of competitors. From igniting the AIGC craze in 2022 to releasing the GPT-5.4 mini lightweight model in 2026, so many years have passed, and its core shortcomings remain unaddressed. Whether you’re a programmer, a self-media creator, or a student, almost everyone I know has had the experience of “using ChatGPT as a last resort, only to be frustrated.” Whether it’s actually good, worth the money, or inferior to domestic models, many people have been misled by the hype. In this review, I will be impartial and unbiased, offering a straightforward user-centric assessment of ChatGPT (equipped with the latest GPT-5.4 series models) from four core dimensions: interface, functionality, competitors, and price. I will lay bare all the pitfalls and truths, providing a user guide that you can directly apply.

Software Basic Information: Named ChatGPT (full name Chat Generative Pre-trained Transformer), developed by the American company OpenAI, it’s essentially an AI tool. It aims to be a jack-of-all-trades, mastering chat, writing, coding, and image drawing, but it doesn’t excel at anything.

Review Background: The AI market in 2026 was incredibly competitive. ByteDance’s Doubao excelled in Chinese language scenarios, Anthropic’s Claude 4.0 was legendary for long text processing, and Google’s Gemini 3 Pro was exceptional for drawing. Only ChatGPT was relying on its past successes. As an “industry pioneer,” the key questions many people have about whether to use ChatGPT are truly viable, whether paying for a membership is a waste of money, and whether it can fix its infuriating flaws (such as nonsensical statements and cross-platform lag).

Core Highlights: To be fair, it can do almost anything. Chatting, copywriting, coding, and image creation are all seamlessly integrated without needing to switch tools. Cross-platform synchronization is relatively smooth. The free version is sufficient for everyday emergencies, and while the paid version isn’t expensive compared to other overseas models, that’s about it.

First Impression: The minimalist interface is an advantage, but it also reveals a lack of effort.

As someone who deals with AI tools daily, ChatGPT’s interface initially struck me as “easy to use,” but a closer look reveals a sense of “perfunctory” design. The lack of unnecessary ads and complex buttons, along with its pure minimalist style, while appealing to beginners compared to the flashy designs of similar models, also suggests a lack of attention to detail.

Visual Design: The overall color scheme is a low-saturation white and gray, accented with a few blue buttons. There are no unnecessary icons or decorations; the interface has only two core elements: the chat window and the function menu. The icons are simple: a new chat is represented by a “+” sign, chat history by a clock, and settings by a gear. Even first-time users can easily understand them. It’s worth praising that the latest version allows switching between light and dark modes. Using dark mode when working late into the night reduces eye strain, which is a thoughtful touch.

Ease of Use: Almost no learning curve. The menu logic is clear: the left side is the chat history, the middle is the chat window, and the right side is the function bar, allowing you to switch models, adjust parameters, and enable image/text features. There are no cluttered hierarchies. The first login includes a three-step tutorial. My friends who had no prior AI experience followed the tutorial and were able to use it in one minute – it’s extremely user-friendly. However, there are also obvious flaws. The “Advanced Parameters” section in the function bar lacks even Chinese annotations. New users seeing terms like “temperature” or “maximum token” will be completely confused and will have to search for them themselves. This is a typical example of “arrogance,” completely disregarding the experience of non-professional users.

Performance: Startup speed is quite good; web browsers open without ads, and the chat interface loads in 2-3 seconds after entering the URL (test device: Lenovo Xiaoxin Pro 14, network: 5G); mobile devices are slightly slower, but still don’t exceed 5 seconds. Switching between conversations and generating content is mostly smooth; even long texts of thousands of words can scroll in real time without freezing. In my actual test, I opened 5 conversation windows simultaneously, letting it code and write copy at the same time, and the interface didn’t lag or crash, which is definitely commendable.

In-depth testing: Seemingly all-encompassing, but riddled with pitfalls
ChatGPT’s biggest selling point is its “all-around capability,” but after testing, it turns out that this so-called “all-around capability” is actually “comprehensive but not refined.” Some functions are so problematic that you might want to uninstall it, especially the text/image function and contextual memory, which are incredibly frustrating.

Text/Image Function: Okay for Writing Text, But a Complete Failure for Drawing

Nowadays, AI is all about integrating text and images, and the latest version of ChatGPT has followed suit by adding an image generation function. The advertising boasts that it can “generate images from text and analyze images in one step,” but in reality, the text generation is barely acceptable, and the drawing function is a complete disaster.

I conducted a real test: Using the theme “The Experiences of Sichuan Farmers in 1647,” I asked it to write a realistic text. Surprisingly, it performed quite well. It not only provided specific people and addresses but also detailed the life scenes from January to December, incorporating the history of the Ming army, the Great Western army, and the Qing army. The details were accurate, and it even suggested adding details about “Huguang Filling Sichuan,” making it highly interactive. However, when I asked it to draw a map of the Sichuan rebellion in 1647 based on the text, it completely failed—the drawing was both crude and absurd. It even depicted the White Banner Army, a force that only existed during the Xianfeng era of the Qing Dynasty, on a 1647 map, without even an “AI-generated” watermark. Unsuspecting viewers could easily mistake it for historical fact; isn’t that misleading?

Regarding the image analysis function, I uploaded a poster with text. It could only recognize the text content, but couldn’t answer questions about the font or color, nor could it extract the poster’s design logic. Compared to the Gemini 3 Pro, it was far inferior—a bronze-level player trying to compete with a master.

Programming and Memory: Useful for programmers in emergencies, but its memory is comparable to Alzheimer’s.

ChatGPT’s programming capabilities have always been touted as phenomenal. The latest version, GPT-5.4, scored slightly higher than Claude Opus 4.1 in professional tests. In practice, it does save programmers some trouble, but it’s not as miraculous as some claim.

I asked it to code a “World War II-themed Huarong Road puzzle game,” and it produced code in 12 seconds. Replacing Rommel with Cao Cao and setting up gameplay where “Allied forces capture German generals,” the code structure was clear, the comments were complete, and it could be run initially by copying it into an editor—a commendable achievement. However, in-depth testing revealed a small bug: the core character “Rommel’s” block was completely immobile, and the game interface was ridiculously rudimentary, requiring manual code modification to function. It was like being given a half-finished product that required rework.

The contextual memory function was even more problematic. It was advertised as remembering key dialogue information, eliminating the need for repeated questions, but in practice, it was akin to Alzheimer’s. For example, I first asked it to introduce the “Transformer neural network architecture,” then followed up with a question about its differences from GPT-3, and it could answer accurately. However, after more than 20 rounds of dialogue, or when discussing multiple complex topics, it completely lost track. Even when I specifically told it “not to use technical jargon,” it still output obscure content, as if it had forgotten what I said before, making me want to close the conversation and reopen it.

Stability and Cross-Platform Compatibility: Decent for Intensive Use, Mobile Version is Purely Superficial

I used it intensively for 3 consecutive hours, simultaneously writing social media content, debugging Python code, and translating English literature. There were no crashes or freezes, and the content generation speed didn’t noticeably slow down, which is acceptable. However, during peak hours (8-10 PM), it’s unreliable. After sending a question, it takes 3-5 seconds for content generation to begin, while during off-peak hours it responds in 1-2 seconds, a significantly worse experience. Additionally, when generating long texts exceeding 5000 characters, there are segmented lags, with each segment taking 1-2 seconds to load, which can be incredibly frustrating when you need it urgently.

ChatGPT supports web, iOS, and Android platforms. Cross-platform synchronization is very smooth; conversations created and parameters adjusted on the web version are synchronized instantly on the mobile version without needing to be reconfigured. The mobile version also supports voice input, and even with slightly accented Mandarin, the recognition accuracy is over 90%. Voice input is quite convenient when typing is inconvenient during commutes.

However, the mobile version is also clearly lacking: the interface is too simplified, and advanced parameters such as “temperature” and “maximum token” cannot be adjusted, so only the default settings can be used; the image generation function is particularly cumbersome to operate, and if you want to modify the image parameters, you have to switch to the web version, which is unnecessary.

[Avoid the Pitfalls Guide]

Don’t Do In-Depth Testing: Features Seem All-Round, But Are Full of Pitfalls
ChatGPT’s biggest selling point is its “all-rounder” nature, but after testing, it turns out that this so-called “all-rounder” is actually “comprehensive but not refined.” Some features are so bad they make you want to uninstall, especially the text/image function and contextual memory, which are incredibly frustrating.

Text/Image Function: Okay for Writing Copy, But a Complete Failure for Drawing

AI is now popularizing integrated text and image functions, and the latest version of ChatGPT has followed suit by adding an image generation function. The advertising boasts that it can “generate images from text and analyze images in one step,” but in practice, text generation is barely acceptable, and drawing is a complete disaster.

I conducted a real-world test: I asked it to write a factual account of “The Experiences of Sichuan Peasants in 1647.” Surprisingly, it performed quite well. It provided specific people and addresses, detailed the daily life from January to December, and incorporated historical information about the Ming, Western, and Qing armies. The details were accurate, and it even suggested adding details about the “Huguang Filling Sichuan” migration, making it highly interactive. However, when I asked it to draw a map of the Sichuan turmoil in 1647 based on the text, it completely failed—the drawing was crude and absurd. It even included the White Banner Army, a force that only existed during the Xianfeng era of the Qing Dynasty, on a 1647 map, without even an “AI-generated” watermark. Unsuspecting viewers could easily mistake it for historical fact; isn’t that misleading?

Furthermore, regarding the image analysis function, I uploaded a poster with text. It could only recognize the text content, but couldn’t answer questions about the font or color, nor could it extract the poster’s design logic. Compared to the Gemini 3 Pro, it’s several levels below; it’s like a bronze-level player trying to compete with a master.

Programming and Memory: A Programmer’s Emergency Solution, But Its Memory is Like Alzheimer’s

ChatGPT’s programming capabilities have always been touted as phenomenal. The latest version, GPT-5.4, even scored slightly higher than Claude Opus 4.1 in professional tests. In practice, it does save programmers some time, but it’s not as miraculous as advertised.

I had it code a “World War II-themed Huarong Road puzzle game.” It produced code in 12 seconds. Replacing Rommel with Cao Cao and setting the gameplay as “Allied forces capturing German generals,” the code structure was clear, the comments were complete, and it could be initially run after copying and pasting into an editor—a commendable achievement. However, deeper testing revealed a small bug—the core character “Rommel’s” block was completely immobile, and the game interface was ridiculously rudimentary. It required manual code modification to function, essentially giving you a half-finished product that you had to rework.

The contextual memory function is even more problematic. Advertised to remember key dialogue information and eliminate the need for repeated questions, in reality, it’s like Alzheimer’s. For example, if I first ask it to introduce the “Transformer neural network architecture” and then follow up with questions about its differences from GPT-3, it can answer accurately. However, after more than 20 rounds of conversation, or when discussing multiple complex topics, it completely gets confused. Even though I specifically told it “not to use technical jargon,” it still outputs obscure content as if it had forgotten what I said before, which is so frustrating that I want to close the conversation and reopen it.

Stability and Cross-Platform Compatibility: Okay for Intensive Use, Purely Superficial on Mobile

I used it intensively for 3 consecutive hours, simultaneously writing social media copy, debugging Python code, and translating English literature, without any crashes or freezes, and the content generation speed didn’t slow down significantly, which is quite reliable. However, during peak hours (8-10 PM), it’s unacceptable. After sending a question, it takes 3-5 seconds for content generation to begin, while during off-peak hours it responds in 1-2 seconds, resulting in a significantly worse experience. Additionally, when generating long texts exceeding 5000 characters, it experiences segmented lag, with each segment taking 1-2 seconds to load, which can be incredibly frustrating when you need it urgently.

ChatGPT supports web, iOS, and Android platforms, and cross-platform synchronization is smooth. Conversations created and parameters adjusted on the web version are synchronized instantly on the mobile version without needing to be reconfigured. The mobile app also supports voice input, achieving over 90% accuracy even with accented Mandarin, which is quite convenient for commuting when typing is inconvenient.

However, the mobile app’s shortcomings are also obvious: the interface is overly simplified, lacking advanced parameters like “temperature” and “maximum token,” forcing users to use default settings; the image generation function is particularly cumbersome, requiring users to switch back to the web version to modify image parameters, which is redundant.

Drawing historical or ID images is prone to errors and lacks AI watermarks, potentially misleading users; even more absurdly, when I uploaded my ID card and asked it to modify the face and information, it executed the modification directly without any warning, posing a significant security risk; 2. Avoid using it during peak hours, especially between 8-10 PM, as the response is too slow and a waste of time; 3. When generating long texts, don’t send questions all at once; send them in segments to avoid lag. Competitive Comparison: Stop Claiming It’s the Ceiling – Domestic Models Have Already Crushed It

With so many AI models available now, ChatGPT is no longer irreplaceable. I’ll compare it to three of the hottest competitors in 2026 – ByteDance’s Doubao, Anthropic Claude 4.0, and Google Gemini 3 Pro – in a straightforward comparison. I’ll skip the technical jargon and focus on the differences that are perceptible to the average user. The conclusion is clear: in some scenarios, it’s not even a fraction as good as domestic models.

Comparable Products: Doubao (domestic ceiling, overwhelming advantage in Chinese scenarios), Claude 4.0 (excellent with long text and programming), Gemini 3 Pro (godlike in drawing and image analysis).

Key advantages:

Versatile and well-rounded: Unlike Doubao, which focuses solely on Chinese language processing, or Gemini, which focuses solely on drawing, ChatGPT can handle chatting, copywriting, coding, and drawing. It’s ideal for those who don’t want to switch between tools, providing a one-stop solution for basic needs.
User-friendly programming experience: The generated code is fully commented, has high error tolerance, supports multiple programming languages, and is easy for novice programmers to understand and modify. Its code is more concise than Doubao’s and easier to debug than Gemini’s.
Better cross-platform synchronization than Claude: Claude 4.0’s mobile synchronization takes 10-15 seconds, while ChatGPT has almost no delay. The web and mobile operation logic is the same, eliminating the need for adaptation.
Lower API price: For enterprise users using it in bulk, its API price is cheaper than GPT-4o and Gemini 2.5 Pro, only one-fifteenth the price of Claude Opus 4.1, saving costs. Crucial Disadvantages (Speaking the Truth Without Mercy):
Completely Outclassed by Doubao in Chinese-Speaking Contexts: This is its most fatal weakness. Its understanding of Chinese contexts and localized expressions is far inferior to Doubao’s. Inputting “interpretation of internet memes” or “Chinese blessings” yields stiff and awkward responses, even failing to understand many dialects and slang terms; Doubao, on the other hand, accurately captures the needs of Chinese users, speaks in a down-to-earth manner, and can even mimic our joking tone. ChatGPT is a complete novice in this regard.
Terrified of Gemini in Image and Text Functionality: Its drawing is rough and prone to errors, and its image analysis can’t even recognize fonts and colors; Gemini 3 Pro produces detailed and accurate images, and its image analysis can extract all details. The difference is immediately apparent when the two are placed side-by-side.
Claude 4.0 is inferior to ChatGPT for long texts: Although it advertises support for 256K context, it often suffers from logical gaps and reads incoherently when handling documents longer than 200 pages. Claude 4.0, on the other hand, handles long texts and academic papers with not only coherence but also precise extraction of key points, far surpassing ChatGPT.
A tendency to spout nonsense with a straight face: This problem persists for years, especially when answering questions about history and professional knowledge. It frequently outputs incorrect information; for example, it previously confused the dates of the White Army while drawing a map, and then described it in a seemingly plausible manner, easily misleading uninformed readers. Doubao, however, has a much higher accuracy rate in localized knowledge and almost never makes such basic errors.

Pricing and Payment Model: Free is sufficient; be cautious about subscribing to paid services—don’t waste your money.
ChatGPT uses a “free + subscription” model, with no one-time purchase. Different versions differ significantly. In practice, the free version is sufficient for ordinary users, while the paid version is only suitable for specific groups. Blindly subscribing is simply wasting your money.

Pricing Structure:

Free Version: No registration required (available from 2024), uses the GPT-5.4 mini model, allows chatting, simple coding, and drawing a few graphs, but has many limitations—limited to 10 chats per day, no more than 1000 characters of text generated per chat, and a maximum of 3 graphs per day. Advanced parameters, long text, and voice input are unavailable.
Subscription: Plus version $19.99 per month, uses the GPT-5.4 full model, has no chat limit or text length limit, all features are available, and priority access to the latest model; Pro version $39.99 per month, adds enterprise-level services and bulk API access, but is largely unnecessary for ordinary users.
API Pricing: Charged per token, $1.25 for 1 million tokens input, $10 for 1 million tokens output, suitable for enterprises to integrate into their products in bulk.

Cost-Effectiveness Analysis (Honestly):

Regular Users (Students, New Graduates): The free version is perfectly adequate for daily copywriting, asking questions, and simple translation. There’s no need to spend money; it’s a waste.
Professional Users (Programmers, Self-Media Professionals, Researchers): The Plus version costs $19.99 per month (approximately 140 RMB), cheaper than Claude 4.0. It saves time on programming and graphic design, making it suitable for long-term use. However, if you primarily use Chinese, don’t choose it; the free version of Doubao offers more features and is far more cost-effective.
Enterprise Users: The API is inexpensive and suitable for batch use, but be aware of its tendency to misinterpret information. Don’t use it to output professional knowledge, or you’ll likely encounter problems.

[Avoidance Guide] Don’t blindly rush for the Pro version! The Plus version is sufficient for regular professional users. The Pro version’s enterprise-level services are completely useless for individuals and a pure waste of money. The free version’s 10 daily usage limits are enough for most people’s daily needs; there’s really no need to spend extra money.

Summary and Recommendations: Choose According to Your Needs, Don’t Be Misled by Advertising

After a week of in-depth testing, my evaluation of ChatGPT (latest 2026 version) is straightforward: it’s not bad, but it definitely doesn’t deserve the title of “AI ceiling.” It’s balanced but mediocre, with many pitfalls, but also some advantages. The key is to consider your usage scenario; don’t be misled by online hype.

Pros and Cons List (Red and Black List, Unbiased)

Red List (Aspects Worth Praising):

Minimalist interface, easy to learn, even beginners can quickly master it;
Balanced overall capabilities, no need to switch between tools, one-stop solution for basic needs;
Smooth cross-platform synchronization, voice input is useful, very practical for commuting or when typing is inconvenient;
The free version has sufficient functions, and the paid version is not expensive compared to overseas models;
Good programming capabilities, saving programmers a lot of rework time.

Blacklist (Frustrating Aspects):

Terribly poor Chinese language support, easily outclassed by Douban (a popular Chinese video platform), with awkwardly stilted localization.
Graphics are frequently flawed and prone to errors; lacks AI watermarks, posing a risk of misleading users.
Prone to making nonsensical statements with a straight face; professional knowledge and historical information are easily misunderstood—don’t believe them.
The mobile app is ridiculously poorly designed; advanced parameters cannot be adjusted, and operation is cumbersome.
Slow response during peak hours; long text generation is laggy, causing significant frustration when needed urgently.

Who should use it, and who shouldn’t:

Best suited for: Programmers (the programming assistance is really useful), people who frequently use AI across platforms, and enterprise users who need to call APIs in batches;
Can be used: Students and working professionals; the free version is sufficient for daily emergencies;
Avoid: People who mainly use it in Chinese scenarios (writing Chinese-style copy, researching local knowledge) should prioritize Doubao; those who want to draw diagrams and analyze images should prioritize Gemini 3 Pro; those who frequently process long texts and academic papers should prioritize Claude 4.0.

Final rating: 3.5 stars (out of 5)

Reason for rating: Not bad, but its shortcomings are fatal. Its Chinese language and image/text functions are crushed by competitors, and it tends to make irrelevant statements. Although it’s well-rounded and good for programming, it doesn’t live up to its current hype. Suitable for people with specific needs; ordinary people should be cautious about buying the paid version—don’t waste your money.

Download/Purchase Recommendations:

Regular Users: Use the free web version directly. No registration is required, and it fully meets your daily needs. Don’t spend money.
Professional Users: The Plus version is sufficient. Don’t bother with the Pro version; it’s a waste.
Enterprise Users: Prioritize the API service. It’s low-cost, but be careful to avoid its misleading claims.
Download Channels: Search for the OpenAI official website on the web. On mobile devices, search for “ChatGPT” in the App Store or app markets. Don’t download pirated versions; they are unsafe.