OpenAI o1: The Reasoning Model That Thinks Before It Answers

What Makes o1 Different

OpenAI's o1, released September 9, 2025, represents a fundamental architectural departure from GPT-4o. While GPT-4o generates responses token by token in a single forward pass, o1 uses a two-phase process: a thinking phase where the model generates an internal chain of thought (not fully visible to users), followed by a response phase where it synthesises its reasoning into a final answer. The thinking phase can run for tens of seconds to several minutes depending on problem complexity.

This approach — inspired by techniques from AlphaGo and reinforcement learning research — was trained with process reward models that evaluated the quality of reasoning steps, not just the correctness of final answers. The result was a model that approached problems more like a careful human expert than an intuitive pattern-matcher.

Benchmark Performance

On AIME 2024 (one of the hardest US high school math competitions), o1 scored 83.3% — placing in the top 500 students worldwide if it had sat the exam. On GPQA Diamond (graduate-level science questions designed to be difficult for PhDs in adjacent fields), o1 scored 78.0% versus GPT-4o's 53.6%. On Codeforces competitive programming problems, o1 reached the 89th percentile of human competitive programmers.

Where o1 Falls Short

Despite its reasoning strengths, o1 had notable limitations at launch. It was significantly slower than GPT-4o for routine tasks — answering a simple factual question could take 30 seconds while GPT-4o answered in under a second. It was also initially more expensive: $15/million input tokens and $60/million output. And its extended thinking made it worse, not better, for tasks requiring quick factual recall or creative writing where deliberate reasoning was counterproductive.

o1 Mini

Alongside o1, OpenAI released o1 Mini — a smaller, faster, cheaper reasoning model at $3/million input tokens. o1 Mini sacrificed some of the breadth of o1's knowledge to focus its reasoning capability on STEM tasks, making it the recommended model for mathematical and coding reasoning where the task domain was narrow and well-defined.

What This Means for Indian Businesses