OpenAI announced o3 and o4-mini models on April 16, 2025, with integrated tool use (web search, code execution, image generation) executed within the reasoning chain, and multi-modal visual reasoning. o3 reports SOTA on Codeforces, SWE-bench (without custom scaffolds), and MMMU; o4-mini achieved 99.5% pass@1 on AIME 2025 with Python access and outperforms o3-mini. OpenAI claims both models show efficiency gains over predecessors at equal cost, with performance continuing to improve under extended inference time. This represents a claimed convergence of specialized reasoning with tool orchestration and signals incremental progress on agentic autonomy and multi-tool reasoning capability assertions.
313 other tracked developments touching the same claims.