Chatgpt becomes a super-KI agent: Openai's new AI models O3 and O4-Mini think now!

Published on: April 17, 2025 / update from: April 17, 2025 - Author: Konrad Wolfenstein

Chatgpt becomes a super-KI agent: Openai's new AI models O3 and O4-Mini think now! - Image: Xpert.digital

More intelligent than ever: Openaai's O series impressed with new skills

Mathematics, programming & more: Openai's O4-Mini is the new AI wonder child! - Openai's O3 understands pictures & solves problems like never before!

On April 16, 2025, Openai presented two new AI models in his O series-O3 and O4-Mini. These are referred to as the most intelligent and most powerful models of the company. The new systems are characterized by improved thinking skills and can use and combine all tools available in chatt for the first time. They were specially trained to think about longer before the answer generation, which makes them particularly effective in complex tasks such as programming, mathematics and visual analysis.

Suitable for:

Chatgpt gets memory and now remembers everything (almost): the new memory function in detail

The new O-series models at a glance

Basic properties and skills

The O-series from Openai represents a paradigm shift in AI development. The models were trained using Reinforcement Learning in order to carry out longer processes of thinking before the answer generation. This approach enables the models to try different solution strategies, recognize errors and to disassemble complex problems into simpler sub -steps.

A significant innovation on O3 and O4-MINI is the ability to use all available chattt tools independently and agent-based. This includes web search, python-based data analysis, image processing, image generation, canvas, automation, file search and memory functions. These tools are integrated directly into the thinking process of the models in order to expand their skills and to manage more complex tasks.

The models can decide for themselves when and how they best use these tools and typically deliver answers in less than one minute, even with more complex problems. This marks an important step towards an agent -based chatt that can carry out tasks independently.

Visual understanding and multimodal skills

A particularly remarkable property of the new models is their ability to “think” with pictures. According to Openaai, this means that you can not only perceive visual data, but can also integrate directly into your thinking process. The models can understand and analyze uploaded images such as whiteboards, sketches and diagrams, even if they are of less quality.

These multimodal skills go beyond pure image processing. The models can curtail or transform images, combine them with other tools and include them in their train of thought to draw well -founded conclusions. This integration of visual data into the thinking process represents significant progress compared to previous AI models.

Performance and benchmarks

O3 as a flagship model

Openai O3 is described as the company's most powerful Reasoning model, which sets new standards in areas such as programming, mathematics, natural sciences and visual perception. In evaluations by external experts, O3 makes about 20 percent less serious mistakes than its predecessor O1 in complex, real tasks.

In various benchmarks, O3 shows impressive results:

It achieves new best values for codeforces and SWE-bench
It sets new standards in the MMMU benchmark for multimodal understanding of understanding
In scientific benchmarks such as GPQA Diamond, which measure questions at PHD level, O3 achieves an accuracy of 87.7% compared to 78% at O1

The model shows special strengths in programming, in the consulting area and in creative tasks. Early testers emphasized his analytical strict as a thinking partner and emphasized his ability to generate and critically evaluate new hypotheses - especially in biological, mathematical and technical contexts.

O4-mini as a cost-efficient alternative

The O4-Mini is a smaller model that has been optimized for quick and cost-efficient processing. Despite its lower size, it achieves remarkable achievements, especially in the areas of mathematics, programming and visual tasks.

It is the most powerful model in the Aime 2024 and 2025 benchmark. In the Aime 2025, it even reached an impressive accuracy of 99.5 percent with access to a Python interpreter. In expert evaluations, it also exceeds its predecessor O3-Mini in non-mint areas and in data science.

Thanks to its efficiency, O4-Mini supports significantly higher usage limits than O3, which makes it a strong option for applications with high volume and throughput that benefit from logical thinking.

Areas of application and availability

Possible uses

With their improved skills, the new models open up a variety of applications:

Complex problem solutions in science and technology, where their ability to disassemble problems into partial steps is particularly valuable
Programming tasks and software development, where you can support the codegenization and troubleshooting
Mathematical and scientific analyzes at a high level
Visual analysis of diagrams, graphics and pictures
Agent -based applications in which the AI independently uses different tools to solve tasks

Availability for users

The new models are gradually made available for different user groups:

Chatgpt Plus, Pro and team users have access to O3, O4-Mini and O4-Mini-High in the model selector since April 16, 2025, where
Chatgpt Enterprise and EDU user receive access within a week after publication
Free users can try out O4-Mini by selecting “Think” in the composer before sending your request
The rate limits for all plans remain unchanged compared to the previous models

Suitable for:

Current developments in Chatgpt von Openaai (March 2025)

Security aspects and further development

Security and robustness: A look behind Openai's new models

Openai emphasizes that both models have been subjected to extensive security tests - according to the company, it is the most comprehensive security program so far. The progressive reasoning skills of the models offer new ways to improve security and robustness. In particular, the models can think about the security guidelines of Openai if they react to potentially unsafe inquiries - a concept called “deliberative alignment”.

The publication takes place under version 2 of the “Preparedness Framework” by Openaai. The company's Safety Advisory Group (SAG) checked the results of the preparedness evaluations and came to the conclusion that O3 and O4-Mini in no of the three monitored categories (biological and chemical skills, Cyberproof and AI self-improvement) achieve the threshold “high”.

Meaning for the AI landscape

The introduction of O3 and O4-Mini is a significant step in the evolution of AI systems. With their improved ability to logically think and integrating different tools, these models approach an agent-based system that can independently solve complex tasks.

With these models, Openaai continues to position itself at the head of AI development, which is also underlined by the recent round of financing, which the company rated $ 300 billion. The combination of improved correcting, tool integration and multimodal skills could significantly expand the area of application of AI and open up new fields of application.

O3 and O4-Mini: Powerful AI models for complex challenges

With O3 and O4-MINI, Openai has presented new AI models, which, thanks to their improved reaction capabilities and the integration of various tools, are significant progress in AI development. The models are characterized by their ability to think through complex problems and use various tools to find solution. While O3 is positioned as a flagship model for demanding tasks, O4-Mini offers a cost-efficient alternative that, despite its lower size, achieves impressive performance.

The new models are already available for various chatters user groups and could expand the spectrum of AI applications thanks to their improved skills. At the same time, Openai emphasizes the importance of security aspects and has subjected the models to extensive tests to minimize potential risks. The development of O3 and O4-Mini marks an important step towards agent-based AI systems, which can increasingly master complex tasks independently.

Suitable for: