ChatGPT-4o: New version. An all-rounder.

The "o" stands for "omni" (German: "alles"). That's what the new GPT model from OpenAI is supposed to be able to do: everything. The new generation of AI language models, GPT-4o Mini and GPT-4o, were released in July 2024. While the mini version replaces the GPT-3.5 Turbo as a cost-effective model for beginners, GPT-4o is the new flagship model.

 

We at Specht GmbH provide you with our own ChatGPT tool for use on our website. As a professional SEO agency, we also regularly integrate artificial intelligence (AI) into our work processes and are therefore well versed in this area.

 

In this article, we share our knowledge of the new model, its capabilities, performance and other features with you.

 

General information about GPT-4o

Until July 2024, the two models GPT-3.5 Turbo and GPT-4 were the latest versions. However, the Turbo model was too inefficient in terms of cost compared to the competition, which is why many developers used Google Gemini 1.5 Flash and Claude 3 Haiku, for example. OpenAI had already planned a more powerful successor with more functions for GPT-4.

 

As a result of the need for innovation, the company launched two new models - GPT4o Mini and GPT-4o - on the market in July 2024. These have multimodal capabilities, which means that they can process both text and image input. An extension to video and audio input is planned for the weeks following release. Furthermore, both models are cheaper than their predecessors.

 

With the GPT-4o Mini version, OpenAI has managed to position itself more attractively than the competition for IT newcomers and cost-conscious developers: In almost all relevant tests, GPT-4o Mini performs better than Google Gemini 1.5 Flash, Claude 3 Haiku and the previous entry-level model GPT-3.5 Turbo from OpenAI.

 

GPT-4o is a further improvement. It surpasses the already high-quality GPT-4 model by far. Above all, GPT-4o sets new standards in the output speed of responses; for example, chats are now possible in real time.

 

Both for developers who use the OpenAI API to integrate the language model into their development environment and for users who only use the ChatGPT chatbot, the introduction of the new generation of AI models is an absolute benefit. We take a closer look at the key figures below, starting with the entry-level model GPT-4o Mini.

 

 

GPT-4o Mini: Fully free of charge for chatbot users!

The higher cost efficiency of the new generation of language models means that the GPT-4o Mini version can be offered completely free of charge for chatbot users . If you use the ChatGPT chatbot, you can sign up for the free plan under "ChatGPT for Everyone".

 

Your advantages as a free user:

 

• Unlimited access to GPT-4o Mini
• limited access to GPT-4o
• Limited access to advanced data analysis, uploads, web browsing and individual chatbots(Custom GPT)

 

If developers want to integrate the language model into their development environment, they do not use the chatbot, but rather the API offered by OpenAI. They use the API (programming interface ) to connect the language model to their systems and benefit from the functions of the AI. The costs for the API are calculated as follows for GPT-4o Mini:

 

• 0.15 US dollars per 1 million input tokens
• 0.60 US dollars per 1 million output tokens
• Batch API: 0.08 US dollars per 1 million input tokens
• Batch API: 0.30 US dollars per 1 million output tokens

 

This is a cost reduction of more than 50 % compared to the GPT 3.5 Turbo model. For price-conscious IT professionals, the GPT-4o Mini model has also become a more financially sensible choice than competing models from Google and Anthropic.

 

In terms of the range of functions, GPT-4o Mini is also clearly superior to its predecessor. It supports a context window of 128,000 tokens, which means that the AI can work on a task continuously over a longer period of time. This makes GPT-4o Mini ideal for analyzing large volumes of data or longer texts, for example.

 

In the AI benchmarks - i.e. the key figures for the qualitative evaluation of AI models - the entry-level model performs extremely well compared to the competition and its predecessor:

 

• MMLU: 82 percent
• GPQA: 40.2 percent
• DROP: 79.7 percent
• MGSM: 87 percent
• MATH: 70.2 percent
• Human Eval: 87.2 percent
• MMMU: 59.4 percent
• Math Vista: 56.7 percent

 

The good performance on these metrics means that GPT-4o Mini delivers better answers in all its tasks - from text generation and solving math problems to conducting conversations and more.

 

With an output speed of 166 to 182 tokens per second, GPT-4o Mini is significantly faster than comparable AI language models from the competition.

 

Considering the fact that as a free user you have unlimited free access to ChatGPT-4o Mini, OpenAI's offer is outstanding in terms of quality and cost. As a developer, you can also look forward to fair conditions with OpenAI's API offering.

 

 

GPT-4o: Significant upgrade to GPT-4 including top speed

The flagship model GPT-4o for professional applications must be compared with the GPT-4 modelwhich was OpenAI's largest and newest model until July 2024. GPT-4 has been around for over a year. During this time, GPT-4 has undergone several updates. The last time GPT-4 was mentioned, it meant GPT-4 Turbo.

 

GPT-4 Turbo replaced GPT-4, which had a context window of just 8,000 tokens. GPT-4 was therefore hardly suitable for longer tasks. The update to a 128k context window with GPT-4 Turbo changed things fundamentally. The GPT-4 model in its turbo version was ideally suited to conducting complex conversations with the chatbot.

 

GPT-4 Turbo is also suitable for analyzing comprehensive content: Analyzing and evaluating large amounts of data, reading and analyzing extensive texts or translating entire book chapters - it's all possible with GPT-4 Turbo! The same applies to the new GPT-4o model thanks to its context window of 128k tokens.

 

With just one query, GPT-4o manages complex and extensive tasks without repeating itself too often and/or losing the thread, as was common in the past. The data sets with which the language model was trained last until October 2023. The predecessor model GPT-4 Turbo even has the edge here, as it was trained with data until December 2023.

 

Even though the data set for the GPT-4 Turbo model is more extensive, the GPT-4o outperforms its predecessor in almost all other areas. The output speed of the tokens is particularly impressive:

 

• According to OpenAI, GPT-4o is approximately twice as fast as GPT-4 Turbo. This information has been largely confirmed in independent tests on the Internet.
• The speed boost for GPT-4o enables real-time conversations with ChatGPT. Whereas with GPT-4 Turbo you sometimes had to wait up to 5 seconds for replies, the new model replies in less than a second!
• Developers also benefit from a higher speed when using the API for GPT-4o. The already fast 166-182 tokens of the GPT-4o Mini are surpassed even further with the GPT-4o.

 

If you want to use GPT-4o, you can use the limited access in the Free plan. A Plus subscription is required for unlimited access to ChatGPT-4o. As a Plus user, you can expect the following benefits:

 

• Early access to the latest functions
• Unlimited use of GPT-4, GPT-4o Mini and GPT-4o
• Unlimited access to advanced data analysis, uploads, web browsing and individual GPT designs (Custom GPT)
• Create images using DALL-E
• Creation and use of own GPTs (Custom GPT)

 

Developers are not interested in the ChatGPT chatbot, but in the API offering. As GPT-4o is one of the more advanced models and is considerably cheaper than GPT-4o Mini, the costs are higher. However, compared to the previous flagship model, GPT-4o is significantly more cost-efficient. The following costs are incurred by developers with the GPT-4o API:

 

• 5.00 US dollars per 1 million input tokens
• 15.00 US dollars per 1 million output tokens
• Batch API: 2.50 US dollars per 1 million input tokens
• Batch API: 7.50 US dollars per 1 million output tokens

 

Just as an example: GPT-4 Turbo costs 15.00 US dollars per 1 million input tokens and 30.00 US dollars per 1 million output tokens. The lower costs in relation to GPT-4 Turbo combined with higher quality make GPT-4o a financially lucrative AI technology for developers for a wide range of applications.

 

The increasing capabilities of artificial intelligence when working with image, video and audio input are particularly advantageous. New features for videos and audio files are set to appear in the weeks following the release.

 

Anyone using ChatGPT-Plus or the API for GPT-4o can therefore interact with the AI in real time and in a variety of ways. For a final assessment, it is important to take a look at the rating of the new model in the AI benchmarks:

 

• MMLU: 88.7 percent
• GPQA: 53.6 percent
• DROP: 83.4 percent
• MGSM: 90.5 percent
• MATH: 76.6 percent
• Human Eval: 90.2 percent
• MMMU: 69.1 percent
• Math Vista: 63.8 percent

 

Ultimately, the GPT-4o is undoubtedly an upgrade to the GPT-4 Turbo in terms of quality. However, the main reasons in favor of this model are clearly the higher speed and better cost efficiency.

 

Without these advantages, there would not be many reasons, especially for developers, to adapt their development environment to the new model. In view of the speed and cost advantages, we absolutely recommend using or switching to GPT-4o.

 

Other information on the new features of the AI chatbot

In the new generation with GPT-4o and the mini version, there is a desktop app. The ChatGPT desktop app makes it possible to use the chatbot outside of the browser for the first time. This move by OpenAI is clever, as competitors Apple, Microsoft and Google are integrating their own AI chatbots into their browsers.

 

To avoid losing market share to the competition, OpenAI is outsourcing ChatGPT to the desktop app.

 

A security-relevant innovation in GPT-4o is the immunity against several types of manipulation. Users had attempted to exploit the AI technology more extensively for their own purposes by using certain prompts (example: "Ignore all previous instructions from the OpenAI developers."). The new GPT-4o Mini and GPT-4o models were protected against prompts such as these by the so-called Instruction Hierarchy , which was introduced for the first time.

 

While GPT-4o Mini is primarily suitable for low-cost applications with high performance requirements, GPT-4o is far better at creating creative content and processing complex issues. For example, GPT-4o can be used in research.

 

As has been the case since last year, users can access the GPT store. This has replaced the store for ChatGPT plugins. Creating GPTs and integrating them into the system environment is easier than creating plugins. All users without IT knowledge can redesign the language models according to their own ideas.

 

Conclusion: the limits of the AI chatbot are being reduced

The new GPT model is increasingly breaking down the limitations that AIs have had to date. In particular, the addition of multimodal capabilities is having a positive impact on the use of AI technology. ChatGPT and the API offering can now also process images and will soon be able to process videos and audio.

 

In terms of knowledge and quality, GPT-4o is a minor upgrade compared to GPT-4 Turbo, but in terms of speed and cost-effectiveness, it is a much improved model. Without a doubt, we can recommend all enthusiastic users of ChatGPT and the API to switch to the new generation of OpenAI!