OpenAI launches a pair of AI reasoning models, o3 and o4-mini

Summary

OpenAI announced on Thursday the launch of o3 and o4-mini, new AI reasoning models designed to pause and work through questions before responding. The company calls o3 its most advanced reasoning model ever, outperforming the company’s previous models on tests measuring math, coding, reasoning, science, and visual understanding capabilities. Meanwhile, o4-mini offers what OpenAI says is a competitive trade-off between price, speed, and performance — three factors developers often consider when choosing an AI model to power their applications. Unlike previous reasoning models, o3 and o4-mini can generate responses using tools in ChatGPT such as web browsing, Python code execution, image processing, and image generation. Starting today, the models, plus a variant of o4-mini called o4-mini-high that spends more time crafting answers to improve its reliability, are available for subscribers to OpenAI’s Pro, Plus, and Team plans. The new models are part of OpenAI’s effort to beat out Google, Meta, xAI, Anthropic, and DeepSeek in the cutthroat global AI race. While OpenAI was first to release an AI reasoning model, o1, competitors quickly followed with versions of their own that match or exceed the performance of OpenAI’s lineup. In fact, reasoning models have begun to dominate the field as AI labs look to eke more performance out of their systems. O3 nearly wasn’t released in ChatGPT. OpenAI CEO Sam Altman signaled in February that the company intended to devote more resources to a sophisticated alternative that incorporated o3’s technology. But competitive pressure seemingly spurred OpenAI to reverse course in the end. OpenAI says that o3 achieves state-of-the-art performance on benchmarks including Codeforces, SWE-bench (without building a custom model-specific scaffold), and MMMU. The company also claims that o3 makes 20% fewer errors than o1 on real-world tasks. OpenAI claims that o3 and o4-mini are its first models that can “think with images.” In practice, users ca...