OpenAI o3
AI Model
Reasoning
Prior generation
The model that broke ARC-AGI and made 'agentic tool use' real.
You're looking at an AI model record: who built it, what it can do, and the hardware and data behind it.
o3 (announced December 2024, released April 2025) scored 87.5% on the ARC-AGI-1 reasoning benchmark in high-compute mode — a test designed to resist memorization — and shipped with full tool use inside its chain of thought: searching, running code and reading files mid-reasoning. Folded into GPT-5's router in August 2025.
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Developed by
Benchmarked on
Succeeded by
Successor to
Related
Sources
Record reviewed as of 2026-01-15 — see Methodology.