o3

o3 is OpenAI's advanced reasoning model, released in April 2025 as the successor to the o1 series . It is designed for complex problem-solving across coding, mathematics, science, and visual perception, with a standout capability that 十字路口Crossing described as "thinking with images" — the model can manipulate, rotate, and zoom images as part of its reasoning chain, blending visual and textual reasoning into a single process . In benchmark terms, o3 outperformed o1 by roughly 20% on tests like AIME 2025 and GPQA, while actually lowering inference costs at equivalent performance tiers . It also marked a shift toward autonomous tool use: the model reasons not just how to use tools like Python or image generators, but when to deploy them .

The model's trajectory has been somewhat unusual. OpenAI first announced o3 in December 2024 during its 12-day release event, then briefly shelved plans for a standalone launch, intending to fold its technology into GPT-5 . The April 2025 release reversed that — o3 arrived alongside o4-mini, with o3-pro promised within weeks . ZhenFund's Yusen Dai, writing in mid-2025, judged o3 a "clear step up" from o1 in practical use, particularly for reasoning length and logical coherence, though he noted it remained an evolution within the Transformer paradigm rather than a paradigm break . Sam Altman indicated o3 and o4-mini would likely be OpenAI's final independent reasoning models before GPT-5's unification .