In September 2024, OpenAI released o1, a new family of language models that approached reasoning differently from anything that had come before in production AI. The models would, before producing their final answer, generate a long internal chain of thought, evaluating alternatives, checking their own work, and deliberating in ways that prior models did not consistently do. The trade was straightforward. The model was slower, often substantially slower, but the answers it produced on hard reasoning problems were substantially better.
The capability shift was particularly visible on tasks that had been difficult for previous models. Complex mathematics, multi-step logical reasoning, and code generation that required careful design rather than recall of common patterns were all areas where o1 outperformed GPT-4o by significant margins. The benchmarks on competitive mathematics and physics problems showed levels of capability that would have seemed implausible only a year earlier.
The model also exposed something interesting about how the broader category of AI systems might evolve. Until o1, the dominant pattern for getting better results from language models was to scale them up. Larger models trained on more data with more compute generally produced better outputs. o1 represented a different scaling axis. Allowing the model to spend more time per query, in a structured way that improved reasoning, was a way to get more capability without making the model larger. The implications for the cost structure of frontier AI were significant if the pattern held.
The product implications were complicated. For applications where latency mattered, like conversational interfaces, o1 was not a drop-in replacement for faster models. For applications where the quality of the output mattered more than the time taken to produce it, like code generation or research support, the tradeoff often favoured o1 strongly.
The release did not fully replace previous models. OpenAI continued to maintain GPT-4o as the default for general use. o1 was positioned as the model for hard problems that justified its slower response time. The product strategy of having multiple models with different capability and speed profiles, and choosing among them based on the task, reflected a maturing understanding of how AI capability would actually be deployed.
What o1 also did was open the door for competitors. Other model providers would release their own reasoning-focused models in the months that followed. The category was now established. The pattern of trading time for capability was real, and would continue to develop through 2025 and beyond.