Bank CEOs meet with Trump to discuss Fannie Mae and Freddie Mac - Bloomberg
Investing.com -- OpenAI today unveiled two new AI models, o3 and o4-mini, which the company says represent its most intelligent and capable reasoning systems to date. Both models are designed to think longer before responding, offering significantly improved performance across complex research, coding, and visual tasks.
For the first time, models in the o-series can autonomously use all ChatGPT tools: searching the web, analyzing data, interpreting images, and generating responses in thoughtful formats. This marks a step closer to agentic AI systems that can complete multi-step tasks independently.
Following the announcement, OpenAI CEO Sam Altman offered a statement on X, saying, "o3 and o4-mini are out! they are very capable. o4-mini is a ridiculously good deal for the price. They can use and combine every tool within ChatGPT. Multimodal understanding is particularly impressive."
The larger of the two models, o3, outperforms predecessors on academic and coding benchmarks such as Codeforces and SWE-bench, and reduces major errors by 20% compared to OpenAI o1. The smaller o4-mini is optimized for speed and efficiency, achieving 99.5% accuracy on AIME 2025 math tasks when paired with a Python interpreter.
Both models are trained to reason strategically about tool use, improving the quality and reliability of outputs. Users can upload images or diagrams—even low-quality ones—and the models can analyze, manipulate, and use them in generating solutions.
OpenAI has also enhanced model safety, retraining refusal mechanisms to address biosecurity and misuse concerns, and embedding reinforcement learning to better govern tool execution. Evaluations found both o3 and o4-mini remain below key capability thresholds for high-risk categories assessed under OpenAI’s Preparedness Framework.
Developers now have access to both models via Chat Completions and Responses APIs, with the latter offering better performance through reasoning token preservation and soon, built-in tools like code interpreters and search. ChatGPT Plus, Pro, and Team users can now select o3 and o4-mini, replacing o1 and earlier mini versions.
OpenAI also introduced Codex CLI, an experimental command-line coding assistant optimized for models like o3 and o4-mini. The tool, now open-source, allows users to pass sketches or screenshots to their terminal and receive multimodal reasoning in return.
Looking ahead, OpenAI says today’s releases point to a convergence between the o-series’ analytical strengths and the natural conversation and flexibility of the GPT line. As AI systems evolve from static responders into dynamic problem-solvers, models like o3 and o4-mini mark a meaningful step toward more capable, autonomous tools that can reason, act, and adapt—an essential milestone on the path to general-purpose AI.