GPT-4 (neural network)

History

2025

OpenAI unveils GPT-4.5 (Orion) with record value of 1 million output tokens of $150

The first thing that should be noted right away is the price. The cost of 1 million output tokens reached an industry record $150 vs $10 for GPT-4o, $60 for thinking o1 and only $4.4 for o3-mini, and for competitors:

• Claude 3.7 Sonnet – $15,
• Qwen2.5 Max – $6.4,
• DeepSeek R1 – $2.19,
• Gemini 1.5 Pro – $5,
• Gemini 2.0 Flash is only $0.4.

The prohibitive price for input tokens is $75, while the GPT-4o has 30 times less - $2.5, o3-mini - $1.1, and o1 - $15, in comparison with competitors GPT-4.5 loses a lot:

• Claude 3.7 Sonnet – $3,
• Qwen2.5 Max – $1.6,
• DeepSeek R1 – $0.55,
• Gemini 1.5 Pro – $1.25,
• Gemini 2.0 Flash is almost free - $0.1.

The industry standard for high-quality models is a wide range from $0.5 to $3, but not $75, which is at least 30-50 times higher than normal.

The price makes the model completely useless for any type of commercial implementation for any type of task, Spydell Finance wrote.

Why? Productivity gains are not growing proportionally in line with incredible price increases.

There is immediately a critically low operating speed (even o1 does not work so slowly, which is in the TOP anti-rating in terms of operating speed) and the lack of decisive progress in generation quality.

In general, in a cursory way, the model is confidently better than an outdated GPT-4o for two years, but the progress is not so obvious in comparison with Claude 3.7 Sonnet and a very controversial and ambiguous comparison with Grok 3.

The problem is that high-quality natural (human-generated) data for training are already running out, which leads to the need to use synthetic data (generated by neural networks), having a dramatic effect on the integral quality of data sets, since there are no mechanisms for reliable verification and validation of ultra-large arrays of synthetic data.

Comparing with DeepSeek R1, o3-mini and other "thinking" models is incorrect, because they function according to different principles.

Comparing GPT-4o with GPT-3.5 two years ago showed decisive superiority, but comparing GPT-4.5 with GPT-4o is no longer so vivid, especially against the background of existing models and progress among competitors.

According to preliminary estimates, among the "non-reasonable" models, GPT-4.5 is either in first place or in the group of leaders around Claude 3.7 Sonnet, Grok 3 and Gemini 2.0 Pro, depending on the configuration of the task.

From the first impressions: the GPT-4.5 is more concise (this is not always a plus, rather even a minus, if the task is worth getting to the bottom of the details) and more natural (the language is more similar to the spoken one, which is also a controversial advantage, since the goal of LLM is to get an answer or solve the problem, and not talk).

GPT-4.5 is an example of scaling uncontrolled learning by scaling computation and data, as well as innovation in architecture and optimization. The result is a model that has broader knowledge and a deeper understanding of the world, resulting in reduced hallucinations and greater reliability across a wide range of themes.

In other words, the model is more "erudite," more packed with all sorts of data and knowledge.

GPT-4.5 supports text generation, image processing, file loading, and a canvas tool for collaborative editing, but does not support multimodal features such as voice mode, video, and screen sharing in ChatGPT. The functionality has been cut down so far, there is no access to the network yet in the developer environment.

OpenAI states that GPT-4.5 understands "human emotions" and intentions significantly better, interprets subtle signals or implicit expectations with more nuance, i.e. better developed "emotional intelligence," resulting in more natural communication.

Another main advantage is improving accuracy and reducing hallucinations. In the PersonQA (hallucination test) test, response accuracy increased from 28% (GPT-4o) and 55% (o1) to 78% in GPT-4.5, and hallucination rates decreased from 52% to 19%.

Comparison with competitors shows competitiveness, but not absolute leadership, which means that the experience of use will determine the functionality, ability to fine-tune and stability of the results (reduced hallucinations).

Given the prohibitive price and low speed of operation, the choice is ambiguous.

GPT-4o and other neural networks do not cope with most programming tasks - OpenAI study

Large language models (LLMs) greatly simplify and speed up the writing of program code, but they are not able to independently cope with most programming tasks. This is stated in the OpenAI study, the results of which were published in mid-February 2025. Read more here.

2024: Pixtral Large neural network with a search engine is presented, which is more powerful GPT-4

In mid-November 2024, the French startup Mistral introduced the Pixtral Large neural network, which can compete with GPT-4. The neural network based on the free chatbot Le Chat is capable of generating images, conducting web searches and serving as an interactive "canvas." Read more here.

2023

Turbo GPT-4 launch

On November 6, 2023, OpenAI announced the GPT-4 Turbo, a more powerful, functional and cheaper version of its large language model (LLM). GPT-4 This "super network" has received an updated knowledge base, which contains information about various events on a global scale until April 2023.

GPT-4 Turbo is available in two versions: one is exclusively for text analysis, and the second understands the context of both text and images. The cost of using the new LLM is $0.01 per 1000 input tokens (approximately 750 words). In this case, tokens are fragments of raw text: for example, the word "fantastic" will be divided into "fan," "tas" and "tic," that is, into three tokens. The price of output tokens - those that GPT-4 Turbo generates based on input - is set at $0.03 per 1000. In the case of image processing, the cost depends on their size. So, according to OpenAI, transferring an image 1080×1080 pixels in size to GPT-4 Turbo will cost $0.00765.

OpenAI announces GPT-4 Turbo

We have optimized performance, so we can set prices on Turbo GPT-4 three times lower for input tokens and two times lower for output tokens compared to GPT-4, says OpenAI.

GPT-4 Turbo has an extended context window - 128 thousand tokens, which is four times more than GPT-4. This is the largest contextual window of any commercially available LLM. The volume of 128 thousand tokens corresponds to about 100 thousand words or 300 pages, which is equivalent to approximately the content of the book "Gulliver's Travels" by Jonathan Swift. The new language model, compared to GPT-4, copes better with tasks that require careful adherence to instructions, such as generating certain formats - for example, "always respond in XML."^[1]

An AI system has been developed on the GPT-4 that quickly trains robots to perform tasks better than humans

On October 20, 2023, the research division of Nvidia Research announced the development of an artificial intelligence-based Eureka system designed to train robots in complex skills. As a result, machines are able to perform some actions even better than people. Read more here.

GPT-4 fooled AI-based defenses: model replaces weapons with apples

The scientist from Google demonstrated how the GPT-4 model bypasses the protection of other models, machine learning which emphasizes the importance chat boats as assistant researchers. This became known on August 1, 2023.

To do this, the researcher asked GPT-4 to develop a method of attack and explain how it works.

AI-Guardian was developed by Gong Zhu, Shengzhi Zhang and Kai Chen and introduced in 2023. AI-Guardian was designed to detect modified images that cheat the classifier, and was GPT-4 involved in bypassing this detection.

For example, adding additional graphic elements to the "STOP" sign can be confusing for self-driving cars. This is one example of a malicious image modification that is scanned by artificial intelligence in a car.

Carlini's work provides Python code proposed by the GPT-4 to bypass AI-Guardian protection measures against attacks. GPT-4 generated scripts and explanations for setting up images to deceive the classifier. So, the classifier may think that a photograph of a person with a weapon is a photograph of a person with an apple. The attacks reduce AI-Guardian's resilience from a reported 98% to 8%. The authors of AI-Guardian admitted that the developed bypass method successfully deceives AI-Guardian protection.

To bypass AI-Guardian protection, it was necessary to identify the mask used by AI-Guardian to detect hostile examples, showing models a variety of images that differ in only one pixel. This "brute force" technique, described by Carlini and GPT-4, ultimately allows you to identify the bypass activation function so that you can then create images to bypass it.

Carlini expects the further development of large language models (LLM).

As the calculator has changed the role of mathematicians, significantly simplifying the execution of mechanical calculations, so today's LLM models simplify the solution of programming problems, allowing scientists to spend more time developing interesting research questions, concluded Carlini[2].

Russian online school Skyeng has introduced GPT-4 technology in training

The online English school Skyeng has launched a virtual interlocutor "Keshu" based on the chatbot GPT-4. The company told about this in mid-March 2023. Read more here.

Creating a Product

On March 14, 2023, OpenAI the chatbot developer company ChatGPT unveiled a new version of its AI-based language model, designated GPT-4.

It is reported that GPT-4 is a large multimodal model trained on a huge amount of data that was not only taken from open sources on the Internet, but also licensed by the developer. These are correct and incorrect solutions to mathematical problems, reasoning of various characters, contradictory and consistent statements and much more. To train the neural network, the Microsoft Azure cloud infrastructure was used. According to OpenAI, in many real-world scenarios, GPT-4 demonstrates "human-level performance."

GPT-4 is a large multimodal model trained on a huge amount of data

The new model can take not only text, but also images as input. In a number of tasks, for example, when processing documents with text and photos, diagrams or screenshots, the GPT-4 demonstrates the same capabilities as when entering only text. Image processing functions are currently being tested, and therefore are not available to the general public.

As OpenAI notes, in a relaxed conversation, the difference between GPT-3.5 and GPT-4 can be barely noticeable. Key differences appear when the complexity of the task reaches a certain threshold - the GPT-4 model is more reliable, creative and capable of processing much thinner instructions than GPT-3.5.

At the same time, the developers say, the new model has the same disadvantages as previous versions. In particular, GPT-4 can "hallucinate" (invent facts) and make mistakes in reasoning. Sometimes a neural network can make simple logical errors and accept obvious false statements from the user. Therefore, as OpenAI notes, great care should be taken when using the output of the language model, especially in "high-stakes" tasks^[3]

Notes