OpenAI o3

OpenAI's o3 represents the full expression of the o-series reasoning philosophy—a model trained to think longer before responding, moving beyond the earlier o1 checkpoint with significantly expanded capability. Rather than simply generating fast responses, o3 was designed to deliberate deeply, deciding when and how to use tools to produce detailed, thoughtful answers. The model can agentically search the web, run Python to analyze uploaded files and data, reason about visual inputs, and even generate images—all while deciding autonomously which combination of capabilities best serves each problem. For everyday users and advanced researchers alike, this represents a step change in what AI assistants can accomplish, particularly on complex, multi-step tasks that require sustained reasoning rather than quick generation.

METR's independent evaluation of o3 provided concrete evidence of this reasoning advancement: on their HCAST benchmark for general autonomy, o3 reached 50% time horizons approximately 1.8 times those of Claude 3.7 Sonnet, suggesting it can handle meaningfully longer tasks independently. METR received access to o3 checkpoints from OpenAI along with technical context and internal benchmark data prior to release, lending credibility to these findings. Though OpenAI announced in May 2026 that o3 would be retired from ChatGPT in August as part of their model lifecycle progression toward newer releases, o3 remains available via API and continues to serve developers and researchers who need its particular blend of deep reasoning and tool-using capability.