OpenAI Unveils o1 Reasoning Model for Harder Math, Coding and Science Tasks

OpenAI has introduced o1, a new family of artificial intelligence models built to handle more complex reasoning tasks in math, coding and science. The company says the model is designed to think through difficult problems more carefully than its earlier systems, delivering stronger results on challenging benchmarks and reducing the kinds of errors that can appear in generated answers.

According to OpenAI, o1 uses reinforcement learning to improve step-by-step reasoning, allowing it to work through problems in a more deliberate way before responding. In internal and benchmark testing, the model outperformed GPT-4o on measures such as AIME, a mathematics benchmark, and GPQA, which evaluates graduate-level science knowledge. OpenAI said the approach helps improve accuracy and lowers hallucinations, or confidently stated but incorrect outputs.

The company is initially making o1 available to some users through ChatGPT, with broader access expected in the near future. OpenAI framed the release as a major step toward models that can tackle more demanding tasks in research, engineering and technical problem-solving. The launch also reflects a growing industry focus on AI systems that do more than generate fluent text, instead emphasizing reliability and deeper reasoning on complex questions.