ChatGPT-4: All about the AI chatbot and its successor models

ChatGPT-4 is a chatbot from the company OpenAI, which was launched on March 14 , 2023 and is based on artificial intelligence (AI). The chatbot is based on the AI model called "GPT-4". Among other things, it is possible to chat with the chatbot, give it tasks and receive solutions.

 

In this blog post, we report on the GPT-4 AI model, its areas of application and capabilities. We also discuss the model's performance data and costs. Information on the successor models GPT-4o Mini and GPT-4o is also part of this article.

 

Areas of application and capabilities of GPT-4

GPT-4 is an LLM that can generate and understand texts in human language. As is common for Large Language Models (LLMs), GPT-4 combines a variety of capabilities. As a further development of GPT-3 and GPT-3.5, GPT-4 sets new standards because it provides more precise answers and is more coherent in its reasoning than its predecessors.

 

While with GPT-3 and GPT-3.5 it was sometimes very common for the language model to invent facts or repeat itself several times during longer explanations, this happens very rarely with GPT-4. The OpenAI company itself says: "GPT can solve difficult problems more accurately - thanks to its broader general knowledge and problem-solving capabilities."

 

Regardless of whether you use the chatbot ChatGPT or the API for developers: Given the higher amount of parameters used to train GPT-4 compared to GPT-3 and GPT-3.5, the quality in the application is higher. In addition, GPT-4 can be used in more application areas than its predecessors:

 

• Creation, analysis and translation of texts
• Solving complex tasks from a wide range of areas
• Implementation of creative tasks
• Creation and analysis of images
• Research on various topics
• Conducting human-like chats with ChatGPT
• Mathematics
• Argumentation and problem solving
• Coding (e.g. developing a program; creating a website)

 

 

Actuality of the data

A special feature of the GPT-4, which is not available with its predecessors or its newer successors, the GPT-4o Mini and GPT-4o, is Internet access.

 

For research, text creation and other tasks, GPT-4 accesses the latest content from the internet as required. This means that the AI model is not limited to the data from the OpenAI training when performing tasks.

 

Speaking of training with data: Even without internet access, GPT-4 is the AI language model from OpenAI that has the most up-to-date data set - even compared to the newer GPT-4o Mini and GPT-4o. The training data extends into December 2023, which allows GPT-4 to retrieve more up-to-date knowledge without using the internet.

 

Multimodal skills

GPT-4 is the first AI model from OpenAI that has multimodal capabilities. By "multimodal", we mean that the artificial intelligence processes not only text input, but also input in other formats.

 

For example, the chatbot ChatGPT-4 can process speech for the first time, so that a verbal conversation with ChatGPT is possible.

 

Furthermore, images can be uploaded and analyzed or edited. The newer GPT-4o Mini and GPT-4o models extend these qualities. OpenAI has announced that video input will also be possible with the new models a few weeks after their release in July 2024.

 

Performance data of GPT-4: High context window, moderate speed

When we talk about GPT-4 today, we are referring to the GPT-4 Turbo version. OpenAI has released several versions of GPT-4 over the years. OpenAI lists the different versions on the Models website under the menu item "GPT-4 Turbo and GPT-4". The most important GPT-4 versions will be discussed here:

 

• GPT-4-0314: This is the first version; it was trained with data up to September 2021 and only has a 4k context window. The small context window makes it virtually impossible to perform longer coherent tasks.
• GPT-4-1106-Preview: A preview model that paved the way for the more powerful and intelligent GPT-4 Turbo.
• GPT-4-Turbo-Preview: Last preview model, which already has a considerable context window of 128,000 (128k) tokens and is trained with more up-to-date data until December 2023.
• GPT-4 Turbo: Final model of GPT-4. This model is installed as standard in ChatGPT-4. It has vision functions so that images can be processed as input. Otherwise it is similar to the Turbo preview model.

 

The fact that GPT-4 Turbo is integrated as standard in GPT-4 chatbots today has a positive effect on the quality of the chatbots. The context window of 128,000 tokens favors the use of the chatbot for longer applications. The reason: an extended context window means that the chatbot can have longer conversations without losing the thread.

 

Advantages through extended context window

Developers who do not use the ChatGPT chatbot but use the GPT-4 language model via the API also benefit from the extended context window. In contrast to GPT-3 and GPT-3.5, developers can use GPT-4 for applications that have a longer context.

 

The fact that several follow-up questions can be asked for each input and the artificial intelligence remains in context increases the precision of the answers. Overall, the extended context window has the following advantages for users:

 

• High degree of creativity
• More information per issue than previous models
• Correct solution of comprehensive mathematical tasks
• Comprehensive coding and programming
• Differentiated argumentation and problem solving

 

Point of criticism: Speed

One shortcoming of the AI model is its slower speed. How long it takes for GPT-4 to provide the desired answers depends on the specific task. However, real-time conversations with the chatbot are definitely not possible.

 

While the new GPT-4o Mini and GPT-4o models have a response time of well under a second, the GPT-4 sometimes takes over 2 seconds and occasionally even 5 seconds for a response. The slow speed is not only noticeable in the chat, but also when using the API.

 

Developers who do not access ChatGPT but the GPT-4 API and use the AI model in their development environment must expect longer waiting times. This has a negative impact on productivity, which is an important factor in professional use.

 

In view of the higher speed and lower costs, it is recommended to use the more efficient AI technology of GPT-4o in comparison with GPT-4. If developers are dependent on an AI model for their applications that can access data from the web, then the GPT-4 model is the better choice.

 

 

AI chatbot: Free use of our tool, fee-based with OpenAI

We at Specht GmbH would like to introduce our ChatGPT tool for use free of charge. In our tool you can get an impression of the functions, quality and performance of GPT-4.

 

The company OpenAI only enables the use of GPT-4 with the paid Plus plan. Developers who want to use the API instead of the chatbot have to pay anyway.

 

Access to ChatGPT-4 for 20 US dollars per month at OpenAI

There is a free plan at OpenAI for users of the ChatGPT chatbot, but the GPT-4 voice model is not part of this plan. GPT-4 can only be used with the paid Plus plan, which costs 20 US dollars per month. As a Plus user, you can expect the following benefits:

 

• Unlimited access to GPT-4, GPT-4o Mini and GPT-4o
• Early access to new features
• Generate images using DALL-E
• Access to the GPT store (replaced the store for ChatGPT plugins in April 2024)
• Creation, use and sale of GPTs
• Advanced data analysis
• multimodal capabilities (GPT Vision)
• Document uploads
• Web browsing

 

API offer from OpenAI: costs depend on the intensity of use

The cost structures are different for developers. From the outset, OpenAI calculated the costs for developers according to how intensively the AI is used.

 

Once the AI technology has been integrated into the development environment via the API (Application Programming Interface), the costs are measured per one million input tokens and per one million output tokens.

 

The best and most efficient are the Turbo models of GPT-4, so it's a good thing that all models of GPT-4 Turbo have the same and most affordable price: 10 US dollars per one million input tokens and 30 US dollars per one million output tokens. The prices for the older GPT-4 models, which are still supported by OpenAI, are as follows:

 

• GPT-4: 30 US dollars per one million input tokens and 60 US dollars per 1 million output tokens
• GPT-4 (32k context window): 60 US dollars per one million input tokens and 120 US dollars per 1 million output tokens
• GPT-4-0125-Preview: 10 US dollars per one million input tokens and 30 US dollars per 1 million output tokens
• GPT-4-1106-Preview: 10 US dollars per one million input tokens and 30 US dollars per 1 million output tokens
• GPT-4-Vision-Preview: 10 US dollars per one million input tokens and 30 US dollars per 1 million output tokens

 

This cost overview shows that some of the older versions are weaker than GPT-4 Turbo, but nothing has changed in their price. If you also consider that the new high performance LLM from OpenAI significantly undercuts the costs of GPT-4 Turbo with higher performance, we definitely recommend developers to switch to GPT-4o if possible.

 

You can find more information about GPT-4o in our blog post about GPT-4o. The new GPT-4o Mini AI model is also of interest to developers. It is the inexpensive entry-level model for cost-conscious developers. The GPT-4o Mini replaces the previous entry-level model GPT-3.5 Turbo with higher performance and lower costs.

 

Further information on GPT-4 and the newcomer GPT-4o

GPT-4o has clear advantages over GPT-4, but ultimately both language models have their shortcomings.

 

AI technology, all voice assistants based on it and all AI tools have limits, especially when it comes to the context window. With the new speech models from OpenAI, this is 128k, which can correspond to 300 A4 pages of text, for example. Beyond this context window, the AI loses all quality in text generation, problem solving and other tasks.

 

Furthermore, AI language models occasionally hallucinate. Hallucinations are the invention of facts. In chat with ChatGPT, it often happens that if you ask misleading questions as if they were self-evident and correct, the AI falls for them.

 

An example of hallucinations could be: "Tell me how Martin Luther ended the French Revolution in 1857." Suddenly ChatGPT starts telling how Luther ended the French Revolution, even though he was long dead in 1857.

 

Furthermore, there is always a risk of prompt injections with AI models and AI chatbots. These are instructions in which users want to manipulate the AI for their own purposes, for example by asking it to ignore the instructions of the OpenAI developers.

 

ChatGPT with the GPT-4 language model was susceptible to such and similar manipulations, but with the new generation around GPT-4o Mini and GPT-4o this danger has been significantly reduced.

 

Conclusion: A popular model that will last a long time

A discontinuation of GPT-4, as is foreseeable with GPT-3.5 Turbo, is not to be expected. Although the GPT-4 AI model is not cost-efficient and by no means the fastest, many free ChatGPT tools on the web work with it, numerous developers have it in their development environment and GPT-4 also provides access to data on the web.

 

Ultimately, GPT-4 has its very own advantages and a raison d'être due to its current widespread use. Nevertheless, we recommend using the GPT-4o Mini and GPT-4o models. Newcomers to the world of AI in particular should give preference to the new AIs from OpenAI.