OpenAI has unveiled two new AI reasoning models, named o3 and o4-mini, focused on delivering more accurate and thoughtful responses by taking extra time to analyze questions thoroughly before replying.
The o3 model, described as the company’s most sophisticated reasoning model developed to date, surpasses its predecessors in key benchmarks including mathematics, coding, logical reasoning, science comprehension, and visual interpretation. OpenAI highlighted o3’s superior performance, noting it makes approximately 20% fewer errors in real-world tasks compared to its first-generation reasoning model, o1.
Meanwhile, the o4-mini model aims to strike a balance among cost, speed, and overall performance—critical considerations for developers selecting AI solutions. Additionally, an enhanced variant dubbed “o4-mini-high,” designed to spend extra time formulating highly reliable answers, has also debuted.
Both models introduce significant advancements over previous iterations. For the first time, these models integrate advanced capabilities within ChatGPT, such as web browsing, executing Python scripts, image processing, and image generation. Notably, OpenAI claims that o3 and o4-mini are its initial offerings capable of “thinking with images.” Users can upload images—such as sketches, PDF diagrams, and even low-quality captures—which the models can interpret, manipulate (for example, by zooming or rotating), and integrate into their responses through a distinct “chain-of-thought” analysis step.
Further enhancing their practicality, the new models can execute Python code within the user’s browser via ChatGPT’s interactive Canvas feature and conduct web searches to deliver updated responses on current events.
OpenAI’s latest launch continues its aggressive push against formidable competitors including Google, Meta, xAI, Anthropic, and DeepSeek. Although OpenAI pioneered the AI reasoning model concept with its o1 model, rival organizations swiftly released comparable—and in some cases, superior—platforms, intensifying the competitive landscape substantially.
Interestingly, the o3 model was initially not planned for deployment within ChatGPT. In February, OpenAI CEO Sam Altman indicated an intention to direct resources toward developing a more advanced model incorporating o3 technology. However, ongoing competitive pressures motivated the company to reconsider and ultimately introduce o3 within the current ChatGPT framework.
Alongside their availability through ChatGPT’s Pro, Plus, and Team subscription plans starting immediately, developers will have access to o3, o4-mini, and o4-mini-high through OpenAI’s API endpoints for building customized applications at usage-based pricing. Additionally, OpenAI intends to launch o3-pro, a variant allocating more computing power to generate higher-quality responses, exclusively for ChatGPT Pro subscribers in the coming weeks.
Looking further ahead, Altman suggested that these models—o3 and o4-mini—might represent OpenAI’s final specialist reasoning models before GPT-5 makes its anticipated debut. This next-generation architecture will potentially integrate standard generative models pioneered by platforms such as GPT-4.1 with advanced reasoning models into a unified system.
