OpenAI o1, released in September 2024, represents a fundamentally different approach to LLM capability. Rather than simply predicting the next token, o1 is trained to spend time "thinking" through problems using reinforcement learning on chain-of-thought reasoning.
**Reasoning Quality**
On competitive mathematics benchmarks (AIME), o1 scores at the 89th percentile of human participants — a level no previous LLM had reached. It solves problems that GPT-4o routinely fails on by breaking them into structured reasoning steps.
**Science & Coding**
o1 ranks in the 89th percentile on the USA Biology Olympiad and achieves 83% on competitive programming problems (Codeforces). These are domain-specific results that demonstrate genuine reasoning rather than pattern matching.
**Latency**
The extended thinking process means o1 is slow by LLM standards. Response times of 10-30 seconds for complex problems are common. This makes it unsuitable for interactive applications but appropriate for batch reasoning tasks.
**Limitations**
o1 does not support image generation, function calling, or browsing in the base API. It is a specialist reasoning tool, not a general-purpose assistant.
**Pricing**
At $15 per million input tokens and $60 per million output tokens, o1 is the most expensive model in OpenAI's lineup. Cost management is essential for production use.
**Verdict**
o1 is the right tool for genuinely hard reasoning problems in math, science, and software engineering. For everything else, GPT-4o is faster, cheaper, and equally capable.