Key Takeaways
- OpenAI’s o3 model can think with images and use tools like web browsing and Python to solve complex, multi-step problems.
- o4-mini is optimized for speed and cost-efficiency, outperforming previous models in math, coding, and non-STEM tasks.
Share this article
OpenAI released two new AI models today, o3 and o4-mini, expanding its capabilities in reasoning and visual intelligence.
o3 represents the company’s most advanced reasoning model to date, while o4-mini offers improved performance for math, coding, and visual tasks at lower costs.
o3 is the first model in the o-series capable of independently utilizing all available ChatGPT tools, including web browsing, Python, image understanding, and generation. Both models introduce “thinking with images,” enabling direct integration of visual inputs into their reasoning process.
The o3 model establishes new benchmarks in software engineering, mathematics, and scientific reasoning, surpassing o1 in tasks requiring detailed analysis, hypothesis generation, and visual content interpretation. External testing shows o3 reduces major errors by 20% compared to o1.
o4-mini, optimized for high-throughput performance, ranks first in benchmarks including AIME 2024 and 2025, demonstrating strong accuracy across STEM and non-STEM fields.
OpenAI also released Codex CLI, a local coding agent for running models from the terminal. A $1 million grant program will support developers building with it.
Both models underwent safety testing using OpenAI’s updated Preparedness Framework, with evaluations confirming risk levels below thresholds in biosecurity, cybersecurity, and self-improvement categories.
The models are available today for ChatGPT Plus, Pro, and Team users, replacing o1 and o3-mini. ChatGPT Enterprise and Edu customers will gain access next week. Free-tier users can test o4-mini using the “Think” option before queries.
OpenAI plans to launch o3-pro in the coming weeks, combining o3’s capabilities with complete tool support for advanced reasoning tasks.
Share this article
