OpenAI o1: models that think before they answer (2024)
OpenAI previewed a model line trained to reason through problems step by step before responding, trading speed for accuracy and reviving the idea that thinking longer can pay off.
On September 12, 2024, OpenAI previewed o1, a new line of models designed to spend time thinking before answering. Where a standard large language model tends to produce a reply in one quick pass, o1 works through a problem in a long internal chain of steps first, and only then gives its final answer.
Test-time compute
The core idea is often called test-time compute: letting a model use more computation at the moment you ask it a question, not just during training. Give o1 a hard maths or coding problem and it will, in effect, deliberate, try approaches, check its work, and backtrack, before committing to a response.
This was a notable shift. Much of the previous era, from GPT-3 to GPT-4, framed progress mainly as training bigger models on more data. o1 emphasised a second lever: a model can also get better answers by thinking longer at the moment of use. The behaviour was shaped with reinforcement learning, rewarding chains of thought that reached correct results.
What it was good at, and what it cost
o1 reported strong gains on tasks with clear right answers: competition mathematics, science questions, and programming, exactly the domains where careful step-by-step reasoning helps most. The trade-offs were just as clear. It was slower and more expensive to run, and OpenAI chose to hide the raw internal reasoning from users, showing only summaries, which reignited debate about how inspectable these systems should be.
It was also a reminder that benchmarks are not the whole story: doing well on contest problems does not automatically translate into reliability on messy, open-ended work.
Why it matters
o1 reframed a question the field had mostly set aside: not just “how big is the model?” but “how long should it think?” By showing that deliberate, multi-step reasoning could be trained and paid for at inference time, it opened a new axis of progress beyond raw scale.
It sits at the end of the stretch of history this timeline covers, and it points forward. The next arguments, about reasoning, reliability, cost, and how to evaluate any of it, start here.