ChatGPT-3.5 vs. 4: What’s the Difference?

When OpenAI released ChatGPT-3.5 to the public in late 2022, it sparked excitement and fear surrounding the breakthrough technology. OpenAI released the next version, ChatGPT-4, in March of 2023 with demonstrated improvements in accuracy, security, memory, context windows, and increased functionality to respond to images and voice prompts. 

In this article, you’ll learn more about the improvements OpenAI made between the GPT-3.5 and GPT-4, as well as examples of how it performs more impressively and information on OpenAI’s pricing model. 

Read more: What Is ChatGPT? (and How to Use It)

What is ChatGPT?

ChatGPT is an application created and managed by OpenAI that allows users to interface with GPT, a large language model that powers generative artificial intelligence. Users can submit requests to the app, and the AI model will consult its training data set to return a unique response to the prompt. 

In practice, ChatGPT can help you compose pieces of writing, such as personal correspondence, marketing materials, or webpage content. You can ask ChatGPT to help you write a grocery list, create a meal plan, design a workout program, describe art, summarize books, complete math problems, suggest code, or translate between languages, among many other things. 

ChatGPT, as we mentioned, is the chatbot you interact with. GPT stands for generative pre-trained transformers and is the brains and power behind ChatGPT’s capabilities. GPT-3 and GPT-4 are large language models created by OpenAI, the organization that also runs ChatGPT, and other artificial intelligence models like DALL-E. 

OpenAI previously trained GPT-3.5 on 175 billion parameters [1]. The organization chose not to reveal the specific details of how they trained GPT-4, partially because OpenAI now operates a for-profit arm, and competition in the space is hotter than when GPT-3.5 debuted. Rumors circulated that OpenAI used more than 100 trillion parameters to train GPT-4, but OpenAI CEO Sam Altman strongly denied those rumors. 

Read more: What Is the BERT Model and How Does It Work?

ChatGPT 3.5 vs. 4

ChatGPT-3.5 sparked much debate and enthusiasm over the many things it could do, such as generate text remarkably similar to human-written work, generate code, and solve math problems. ChatGPT-4’s capabilities are so much greater than ChatGPT-3.5 that a Microsoft research team called it an “early (yet still incomplete) version of an artificial general intelligence (AGI) system” [2]. The team pointed out a few potential applications for GPT-4, including in the fields of mathematics, coding, vision, medicine, law, and psychology. 

GPT-4 demonstrates superiority over GPT-3.5, even using human academic achievement standards. OpenAI reported that GPT-3.5 passed the Uniform Bar Exam with a score that would rank in the 10th percentile of test takers compared to actual people aspiring to become lawyers. GPT-4 performed remarkably better, ranking in the 90th percentile compared to humans [3]. Similarly, GPT-3.5 ranked in the 31st percentile compared to high school students competing in the Biology Olympiad, and GPT-4 scored significantly better, scoring in the 99th percentile [3]. 

Regarding the technical differences between GPT-3.5 and GPT-4, the latter offers increased functionality, memory, and safer and more accurate responses. Let’s look at a few specific upgrades when comparing ChatGPT-3.5 versus ChatGPT-4. 

1. ChatGPT-4 is multimodal and can respond to visual and audio input. 

In the previous version, you needed to write a prompt using text to generate an output from ChatGPT. With version 4, you can still use text, but you can also offer an image or even a voice command to make a request from the application. OpenAI’s example of this new feature is that you could put in a picture of the inside of your refrigerator, and ChatGPT-4 could suggest recipes you could make with the ingredients in the image. You can also speak to ChatGPT-4, and the AI will generate a voice to speak to you. Multimodality also allows ChatGPT to handle more functions, like captioning or translating videos. 

2. ChatGPT-4 is safer and more accurate than GPT-3.

According to OpenAI, ChatGPT-4 is “82 percent less likely to respond to requests for disallowed content and 40 percent more likely to produce factual responses than GPT-3.5 on our internal evaluations” [3]. The organization reports that it achieved this new security and accuracy using user feedback, consultations with security experts, and real-world applications. 

This upgraded feature is good news considering the controversies OpenAI is involved with surrounding misinformation, inaccuracy, and bias in its answers. One example of such controversy is a groundbreaking defamation lawsuit where a man claimed ChatGPT falsely reported that he embezzled money.

3. ChatGPT-4 has more memory than GPT-3.5.

ChatGPT-4 can remember a lot more of a conversation with a user. Not only was ChatGPT previously limited to information available online before 2021, but it also had a limited short-term memory of about 8,000 words. In conversation, the program might remember about 8,000 words until it starts forgetting what you discussed previously. In comparison, GPT-4 has a short-term memory eight times larger at 64,000 words. 

ChatGPT-4 also has a longer context window, or the amount of text it can process simultaneously. For example, you could ask ChatGPT-4 to analyze a document for you, and it can now process about 25,000 words at a time. Another version of the technology called ChatGPT-4 Turbo can process up to 128,000 words. With this feature, you could include a website link in your prompt and ask ChatGPT to consider that source when giving its answer. 

ChatGPT can now access the internet.

In other news regarding ChatGPT upgrades, the program can now access the internet in real-time and provide summaries or respond to links. Previously, you could not access the internet in real-time using ChatGPT, and the program was limited to a data set containing information available before 2021. OpenAI released this update on September 27, 2023, allowing ChatGPT users the ability to discuss current events, ask ChatGPT questions about websites, or suggest links to find more information. Currently, this new feature is only available to paid ChatGPT subscribers, but it is not exclusive to GPT-4. 

ChatGPT pricing

ChatGPT-3.5 is available for free to the general public, but to access GPT-4, you will need to subscribe to the service called ChatGPT Plus. With the subscription, which costs $20 per month, you also gain access to additional features, including:

• Plugins: Plugins are tools you can use to connect ChatGPT to other applications and accomplish more tasks. For example, plugins from companies like Expedia, Kayak, and OpenTable make it possible for you to use ChatGPT to make travel arrangements. 

• DALL-E: DALL-E is generative artificial intelligence you can access inside ChatGPT that can create novel images from the prompts you provide. Using the same natural language understanding that provides you with text prompts, DALL-E can create images of real or imagined things in specific styles or any other description you can think of. 

• Personalized GPTs: Personalized GPTs are another feature available to plus users. This tool is a chatbot trained with GPT and customized to fit a specific need. For example, you might customize a GPT to help you study for a class by providing it with your notes and course materials. 

What is ChatGPT-4 used for?

We mentioned above some of the things you can use ChatGPT for, such as solving math problems, writing essays, translating languages, or writing computer code. However, as the Microsoft research team pointed out, this technology has the potential for even greater applications. Here are some of the ways companies are using ChatGPT: 

• Be My Eyes: Danish accessibility app Be My Eyes uses GPT-4 to work on its Virtual Volunteer service to help connect individuals with assistance when human volunteers are unavailable—a technology only possible with the multimodality of GPT-4. Be My Eyes is a tool for people with low vision or who are blind to connect with sighted volunteers remotely in a moment where they could use a small amount of assistance, such as checking the expiration date on a product before they consume it.  While the tool can be very useful, it relies on volunteers to log in and provide simple services without Virtual Volunteer. 

• Morgan Stanley: Financial powerhouse Morgan Stanley uses a custom GPT model trained on the company’s vast library of documentation, data, and analyst insights. This technology allows instant access to a virtual agent with extensive knowledge of investment strategies, wealth management content, and information about financial products. Accessing this information before using a GPT involved searching through folders and opening PDF files to locate the information you were looking for. A customer GPT allows users to request the information they’re looking for using natural language and receive a generated expert answer. 

• Duolingo: The language-learning app Duolingo uses GPT-4 to provide conversational practice for students learning a new language. Duolingo provides engaging lessons for students, but students were previously limited in options to engage in organic conversations. ChatGPT allows users to engage in a natural conversation with a virtual study buddy who will respond in a way that simulates the way a real person might respond.  

Learn more with Coursera.

Consider taking an online course to take the next step and learn more about using generative AI like ChatGPT. Prompt Engineering for ChatGPT offered by Vanderbilt University, is a beginner-level course that can teach you to use ChatGPT more effectively. You can also learn to work with large language patterns and chain of thought prompting.