OpenAI is preparing to launch GPT-5, the successor to GPT-4, but internal sources report only moderate improvements compared to the previous version. Testing of the model showed better performance in programming, mathematics, and complex instructions, such as automating customer inquiries. However, the difference between GPT-5 and GPT-4 is smaller than it was between GPT-3 and GPT-4.
It is known that GPT-5 creates more user-friendly applications and better allocates computational resources. At the same time, the company faced scaling challenges and a lack of quality training data. The model, codenamed “Orion“, did not meet expectations and was released as GPT-4.5, which operated slower and more expensively than GPT-4o.
OpenAI is also working on so-called “reasoning models” that perform complex tasks better, such as solving mathematical problems or searching for information on the internet. Among these models are Q*, o1, and o3, which are based on GPT-4o and use a reinforcement learning approach. However, when o3 was adapted for chat mode, its performance in communication decreased, and the costs for simple responses increased.
GPT-5 is expected to become the foundation for more autonomous AI agents that perform multi-step tasks with minimal human oversight. The model better determines how much resources are needed for each task, which can reduce unnecessary expenses. OpenAI increasingly uses a “universal verifier” that automatically evaluates the quality of the model’s responses even in creative tasks.
Despite limited technical achievements, OpenAI hopes that GPT-5 will help the company maintain customer interest and strengthen its position in the AI field. This is especially true for coding automation, where Claude models from Anthropic currently lead.
