OpenAI o1 Launches: 'Think Before Answering' Opens a New Reasoning-Model Paradigm

OpenAI released o1 (preview; internal codename Strawberry) on September 12, 2024 — its first model built for reasoning: large-scale reinforcement learning trains it to produce a long chain of thought before answering, sharply beating GPT-4o on math, science and code.

o1's significance is less the model than the new scaling axis it opened: test-time compute, trading more inference time for better answers. o3, DeepSeek R1 and Gemini Deep Think all followed this path, making reasoning models the main battleground of 2025.

From pretraining scaling to inference scaling, o1 marked the official opening of the second curve of large-model capability.

Why 'Test-Time Compute' Is a New Scaling Axis#

Before o1, the main road to more capability was pretraining scaling: more data, bigger parameters, more compute. But that curve is costly and hits diminishing returns. The second curve o1 opened: let the model think longer before answering, trading inference compute for answer quality. The key value is that when pretraining meets data and compute ceilings, test-time compute offers a fresh dimension of improvement. o3 then wove tool calls into the chain of thought (see our o3/o4-mini coverage), and DeepSeek R1 used reinforcement learning to pull open reasoning to the frontier (see our coverage) — all advancing along this curve.

Our Take#

o1's real weight is not any single model's score but the new technical route it opened: the 2025 reasoning-model race and the agent-capability race were fundamentally doubling down on this curve. For ordinary users it made 'wait a moment on hard questions, instant answers on easy ones' the new normal; for the industry it shifted the competitive focus partly from 'how big is the model' to 'how clever is the reasoning.' One often-overlooked detail: the rise of reasoning models directly raised inference-compute demand, planting the seed for the later compute-capital wave (see our Stargate and Nvidia coverage).