OpenAI has recently launched o3 and o4-mini, the latest additions to its line of reasoning models, following the introduction of the GPT-4.1 family. The o3 model, previewed in December, is OpenAI's most advanced reasoning model to date, while o4-mini is designed as a smaller, more efficient alternative. Reasoning models are engineered to process prompts with greater deliberation, resulting in higher-quality responses. Both o3 and o4-mini demonstrate strong performance in areas like coding, math, and science, building upon the capabilities of previous models. A significant enhancement is their newfound ability to understand visual information. The o3 and o4-mini models mark OpenAI's first foray into models that can "think with images." This means they don't simply recognize images; they can utilize visual data in their reasoning process. Users can upload images, even those of low quality, and the models will still be able to interpret them effectively. This capability opens up new avenues for AI-driven analysis and problem-solving. Another notable advancement is the ability of o3 and o4-mini to independently utilize ChatGPT tools, including web browsing, Python execution, image understanding, and image generation. This allows them to tackle complex, multi-step problems more effectively. OpenAI suggests this represents a move toward a more agentic ChatGPT, capable of independently executing tasks on behalf of users. The models can now leverage advanced tools to achieve better results, similar to how a person might use a calculator to solve a complex equation. In a demonstration, the o3 model was presented with a scientific research poster and tasked with drawing a conclusion not explicitly stated in the poster. The model autonomously browsed the internet and zoomed into different elements of the image to arrive at a conclusive answer, showcasing its ability to use multiple tools and analyze images in detail. This illustrates the potential for these models to assist in research, analysis, and other complex tasks. According to OpenAI, the o3 and o4-mini models offer improved instruction following and more verifiable responses compared to previous generations. Benchmarks indicate that the new models outperform their predecessors, even without utilizing their expanded toolset. These improvements make them more reliable and useful for a wider range of applications. Reports indicate that the new models can synthesize information from various fields and use that knowledge to suggest innovative experiments. These experiments could cover complex topics such as nuclear fission or pathogen detection. While OpenAI has not directly commented on these specific capabilities, the potential for AI to contribute to scientific discovery is significant. The o3 and o4-mini models are currently available to ChatGPT Plus, Pro, and Team subscribers. They appear in the model picker as o3, o4-mini, and o4-mini-high, replacing older models. Pro users will soon have access to o3-pro. The models are also accessible to developers through the API. To address safety concerns, OpenAI has subjected both models to rigorous stress-testing and evaluation under its Preparedness Framework. Details of these evaluations are available in the system card. These measures aim to ensure the responsible development and deployment of these advanced AI models. In addition to the new models, OpenAI has launched Codex CLI, an open-source coding agent that runs locally in users' terminals. This tool allows users to connect AI models, including o3 and o4-mini, to their own code and tasks. Codex CLI is available on GitHub, providing developers with a valuable resource for integrating AI into their workflows. OpenAI has also announced a $1 million initiative to support early projects, awarding grants in $25,000 increments via API credits. Proposals can be submitted on the OpenAI site. With the release of o3 and o4-mini, OpenAI continues to push the boundaries of AI capabilities. The combination of enhanced reasoning, visual understanding, and agentic tool use positions these models as powerful tools for a wide range of applications, from research and development to everyday problem-solving. As AI technology continues to evolve, these advancements offer a glimpse into the future of human-computer interaction and the potential for AI to augment human capabilities.
