GPT-4 Statistics

OpenAI released its latest multimodal large language model “GPT-4” last year on 14th March 2023. The advanced LLM GPT-4 has been making waves globally ever since its release. This powerful language model offers excellent capabilities such as text generation, image recognition, translation, code writing, and more.

In this article, we are going to dive deep into GPT-4 statistics and understand its processes and trends. We will also include the number of parameters GPT-4 is trained on, the performance of GPT-4 on standardized tests, benchmark comparisons, and more. 

Top GPT-4 Statistics 

• GPT-4 was released by OpenAI on 14 March 2023.
• GPT-4 is capable of processing and generating inputs and outputs in various formats such as text, image, audio, and video. 
• The response time to audio input in GPT-4 can go as little as 232 milliseconds, with an average of 320 milliseconds. 
• GPT-4 is capable of Achieving 88.7% on 0-shot COT MMLU for general knowledge questions.
• The context length of GPT-4 is 8,192 tokens. 
• GPT-4 Turbo is capable of considering about 96,000 words within one go. 
• OpenAI’s GPT-4 can outperform Whisper-v3 in speech recognition and translation.

Overview of GPT-4 

What is GPT-4?

GPT-4 is the latest large language model developed by OpenAI, the organization that created the popular AI chatbot ChatGPT. This LLM is known for its impressive capabilities of text creation, code generation, and translation. 

GPT-4 is a Multimodal LMM 

The stand-out quality of GPT-4 that sets the large language model apart from its predecessors GPT-3 is its multimodality. This means that GPT-4 is capable of processing and responding to data and information in numerous formats such as: 

• Text (written language) 
• Images (visual information) 
• Code (programming language) 
• Audio 

Was GPT-4 Really Trained on 100 Trillion Parameters?

No, GPT-4 wasn’t trained on 100 trillion parameters. OpenAI CEO Sam Altman claimed the rumor is untrue. 

While other sources claim GPT -4 has close to 1.8 trillion parameters across 120 layers, this makes it almost ten times larger than its previous LLM, GPT-3, which had around 175 billion parameters.

ChatGPT Plus Subscribers (GPT-4 User Numbers)

There are around 180.5 million active users on ChatGPT. An estimated 1% of subscribers have subscribed to ChatGPT Plus, which gives them access to GPT-4 for a monthly price of $20. This makes the number of users around 180,000 to 200,000. 

ALSO READ: Number of ChatGPT Users (June 2024)

GPT-4 has a context length of 8,192 tokens 

With a context length of around 8,192 tokens. OpenAI has started to provide limited access to 32,768 content (which is the text of about 50 pages) version GPT-4-32k. Previously, GPT-4 had featured a content window of 8,000 tokens with a 32K model available for some developers via API. However, this context window has now been surpassed for turbo models. Therefore, this indicates that GPT-4 Turbo is capable of considering about 96,000 words in a single go which is higher than novels. 

GPT-4 Test and Exam Performance

Source: OpenAI 

GPT-4 vs GPT-3

In this section, we are going to compare GPT-4 with GPT-3 to understand its true capability and how it performs in terms of generating relevant and accurate answers:

• Accuracy and Relevance: In terms of providing more relevant and precise answers to users’ questions. Studies have found GPT-4 to be a more capable large language model that can generate answers with an accuracy of 85.5% in a 3-shot massive multitask language understating. GPT-3.5 had an accuracy rate of 70.1% which is lower than GPT-4.  
• Image Recognition: OpenAI’s latest LLM GPT-4 can see, analyze, and describe images with complete detail. For example, if you show an image of a particular food item, GPT-4 can instantly suggest recipes for you. GPT-3 on the other hand, doesn’t have an image recognition feature available. 
• Audio and Vision: ChatGPT’s premium plan known as “ChatGPT Plus” allows users to engage in conversation with GPT-4 through its voice control feature which is available on the mobile app. It also allows users to effortlessly explain and describe things to the GPT via drawing. 
• Mathematical Skills: Studies suggest that GPT-4 is much more capable of solving mathematical problems and generating accurate and relevant answers to users’ questions compared to GPT-3. 
• Language: The language test revealed that GPT-4 has outperformed GPT-3.5 in 24 out of 26 tests, including low-resource languages such as Welsh, Latvian, and Swahili. In addition, GPT-4 also outperformed GPT-3.5 in paraphrasing, text edition, and solving grammatical errors.
• Safety Measures: When it comes to privacy and security measures, GPT-4 is better than GPT-3.5 in avoiding sensitive, disallowed, and inappropriate requests. Apparently, OpenAI has decreased its model’s tendency to respond to disallowed content by 82% and improved its response to sensitive requests by 29%.

Benchmark comparison between Google’s Gemini Ultra and OpenAI’s GPT-4V in 2024

A benchmark comparison was made between OpenAI’s GPT-4 and Google’s Gemini Ultra to understand its true capability and performance. Gemini Ultra has beaten OpenAI’s GPT-4 in the MMMU benchmark. The only two categories where GPT-4 performed better than Gemini Ultra were Business and Science, while the overall quality of the models was pretty similar. 

Below we have mentioned a table showcasing the benchmark comparison of Google’s Gemini Ultra and OpenAI’s GPT-4:  

Source: Statista 

Code Benchmark comparison between Gemini Ultra and GPT-4 in 2024

OpenAI’s GPT-4 and Google’s Gemini Ultra are currently some of the most popular generative AI platforms globally, utilized by users worldwide. Surprisingly, both GPT-4 and Gemini Ultra have quite similarities in the code creation benchmark. Gemini Ultra is a bit ahead of GPT-4, beating it in both the coding benchmark and MMLU benchmarking. 

Below we have mentioned a table showcasing a code benchmark comparison between Gemini Ultra and GPT-4 in 2024: 

Source: Statista 

ALSO READ: AI Usage Statistics 2026: 16.3% of the global population uses generative AI

Top Commercial Uses of GPT-4

In this section we are going to take a look at some of the top commercial uses of GPT-4: 

1. Duolingo: A Language Learning 

Popular language learning app Duolingo turned towards OpenAI’s newest large language model GPT-4 to enhance its capabilities by adding two new features: A “Roleplay” feature which acts as an AI conversation partner and A “Explain my Answer” feature that breaks down the rules and alerts the user when they make a mistake along with the reasoning. Users can access both of these features through Duolingo’s new subscription tier known as Duolingo Max. By adding the latest features, users will be able to learn from their mistakes as GPT-4 generates initial responses by simply clicking on “Explain my answer.” From there, users can directly return to the lesson or gain further in-depth explanations. (Source) 

2. Be My Eyes, Visual Impairment Assistant

The Danish mobile app Be My Eyes is a visual impairment assistant specially built to help blind or visually impaired users recognize elements and objects. Be My Eyes collaborated with OpenAI in March 2023 to create a new tool called “Be My AI” that can describe the visual world to blind or visually challenged people. This new product was incorporated with OpenAI’s latest LLM GPT-4V which assists in providing a description of the images taken on the smartphone by the blind user. “Be My AI” was piloted from March 2023 to August 2023 with approximately 200 blind or low-vision beta testers to assess the safety and user experience of the tool and point out if the platform required any room for improvement or changes. By September 2023, the group of beta testers had reached 16,000 blind or visually impaired users who were requesting almost 25,000 descriptions a day. (Source)

Be My Eyes also followed it by the development of a virtual volunteer which is also powered by OpenAI’s GPT-4 available within the app of Be My Eyes that can create the same level of understanding and context as a human volunteer. ?(Source)

3. Stripe, Fraud Detection 

The American multinational financial services and software, Stripe has leveraged OpenAI’s GPT-4 to fight fraud and enhance its user experience. Previously, Stripe has utilized GPT-3 to ensure a good support team and help its users through various tasks such as summarization of a user question and routing issue tickets. But with GPT-4 Stripe is able to perform various additional tasks such as scanning sites and providing a short summary of the available data, which apparently outperformed human-written summaries. 

It was revealed that when performing a hand-checking on the results it was found that humans were wrong and in fact summaries generated by the model were correct. The Product Lead from Stripe stated, “GPT-4 was better compared to human reviewers.”

Another excellent capability of Stripe includes its extensive technical documentation and robust developer support team that can effortlessly answer all the troubleshooting issues or technical questions ?instantly. GPT-4 can even scan the inbound communications and recognize any inappropriate or coordinated activity. 

ALSO READ: AI Coding Tools Statistics, Market Size and Growth 2025-2026

4. Icelandic Government, Language Preservation 

Iceland has collaborated with OpenAI for the utilization of its advanced LLM GPT-4 along with the help of private industry to preserve the Icelandic language and to transform the defensive position into finding new ways to innovate. The core reason behind this collaboration was not only to boost the capability of GPT-4 to offer service to a new corner of the world but also to generate resources that truly promote the preservation of various low-resource languages. 

In the Reinforcement learning from Human Feedback, human testers provided a prompt to GPT-4 and it reverted back with four possible answers to the tester’s input. Once the answers are generated the testers select the best possible answer from the available options and further edit it to generate the best output. The data from this process is then utilized to train GPT-4 to generate improved responses in the future. (Source)

5. Khan Academy, Education

Khan Academy made an announcement in March 2023, that it will utilize GPT-4 to power Khanmigo which is an AI-powered assistant that operates both a virtual tutor for students along a classroom assistant for teachers. (Source) 

The non-profit organization began its testing process with OpenAI’s latest language model back in 2022 and initially started by making the Khanmigo pilot program specially available for a limited number of users, while others can join the waitlist. Apart from this, Khan Academy has also started the adoption of GPT-4 for teachers. It has begun the testing process through which teachers can utilize GPT-4 for different purposes such as writing classroom prompts or generating instructional materials. 

6. Morgan Stanley, Wealth Management 

Morgan Stanley has leveraged GPT-4 to arrange and sort out its extensive knowledge base. 

At the beginning of last year, the company started to explore ways to harness the intellectual capital with GPT’s retrieval and embedding capabilities first with the use of GPT-3 and then GPT-4. ?

Morgan Stanley has utilized GPT-4 to make its internal chatbot more helpful for the company’s needs. Today, over 200 employees are querying the system every day and providing their feedback effortlessly. 

FAQs 

Can I use GPT-4 for free?

Yes, users can use GPT-4 Turbo for free in Microsoft Pilot and receive accurate and better results by accessing the latest AI chatbot capabilities. 

What app uses GPT-4?

Various apps utilize GPT-4, such as the language learning app Duolingo, the Visual Impairment Assistant Be My Eyes, and more. 

Can GPT-4 do coding?

Yes, GPT-4 can perform coding. GPT-4 can even assist developers by generating various useful suggestions, creating code snippets, and offering various insights on coding. It can be utilized as a complement to human coding skills instead of acting as a replacement. 

Can GPT-4 generate images?

Yes, GPT-4 is capable of generating images. It’s a multimodal large language model, and therefore, it has capabilities including text, image, audio, and video outputs. 

Wrapping Up 

GPT-4 is a powerful and versatile language model and its immense scale, safety features, and capability to handle different forms of data have increased its demand among users across the world.

But GPT-4 is still relatively new and with higher accessibility and functionality the influence of the LLM is expected to grow. From content generation to coding to summarization to image recognition GPT-4 has the potential to transform the way we interact with technology and data.