OpenAI o1

OpenAI o1 was the first publicly available frontier reasoning model and the existence proof for spending extra inference-time compute on a private chain of thought before answering. The September 2024 preview release and December 2024 GA established the template: the model produces a hidden reasoning trace (billed at output rates) before the user-visible answer, with benchmark scores on GPQA-Diamond and AIME that materially exceeded GPT-4o on the same architecture-and-data class. The pricing structure was new to the market: at $60 per million output tokens with reasoning traces consuming most of the output budget, a single hard problem could cost dollars rather than fractions of a cent. That created two follow-on stories. First, the open community responded with DeepSeek R1 four months later under MIT license, demonstrating that the reasoning recipe was within reach of organizations not at OpenAI's scale. Second, the reasoning-vs-cost framing made "thinking budget" a first-class deployment knob: subsequent OpenAI releases (o1-mini, o3, o3-mini) and competitor responses (Claude 3.7 extended thinking, Gemini 2.5 thinking) all let the developer dial how much inference compute to spend per request.