o3-mini

o3-mini is a smaller, cheaper OpenAI reasoning model positioned for everyday productive tasks rather than frontier research. In ZhenFund's framing, it sits in a three-tier hierarchy: o3-like models tackle elite-level exploration (scoring 2700 on Codeforces, a level reached by only ~130 humans worldwide), while o3-mini handles practical work "like a programmer," and simpler models answer basic on-device queries such as weather lookups . Yusen Dai of ZhenFund noted that o3-mini achieves strong results through longer inference time—its "o3 mini high" mode performs very well—while reportedly activating only 3.7B parameters per forward pass, a scale he found "shocking" . The model also serves as a testbed for compressing reasoning: as Peak (Ji Yichao) put it, future models may push reasoning compression toward a limit, but it won't disappear entirely, and o3-mini is "precisely exploring this process" . On SWE-bench, a benchmark for AI coding ability, o3-mini scored in the 60-70 range by late 2024 alongside o1, up from GPT-4's 4-5 points at the start of that year .