From prompted reasoning to trained reasoning
A way of writing the examples in a prompt becomes a training method, and within two years every lab offers models that think before they answer.
GPT-3 (May 28, 2020)
GPT-3 shows that a large language model can solve new tasks from a few examples in the prompt.
Chain of thought (January 28, 2022)
Wei and colleagues change only the examples. If the examples include the intermediate reasoning steps, models like GPT-3 and PaLM write out their own steps and get far more answers right.
Tree of thoughts (May 17, 2023)
Tree of Thoughts generalizes the chain. The model explores several paths, evaluates them, and can backtrack.
OpenAI o1 (September 12, 2024)
OpenAI stops asking for reasoning in the prompt and trains o1 with reinforcement learning to think before it answers. More thinking time yields better answers.
DeepSeek-V3 (December 26, 2024)
Meanwhile, in China, DeepSeek releases V3, an open model trained with far less compute than usual.
DeepSeek-R1 (January 20, 2025)
Building on V3, DeepSeek reproduces o1's paradigm with open weights. Its R1-Zero variant learns to reason through reinforcement learning alone.
Grok 3 (February 17, 2025)
xAI adds a mode to Grok 3 that reasons before answering.
Claude 3.7 Sonnet (February 24, 2025)
Anthropic combines quick answers and extended reasoning in a single model, with a thinking budget that can be adjusted.
Gemini 2.5 Pro (March 25, 2025)
With Gemini 2.5, Google makes every model in its new generation think before answering. Reasoning is no longer a separate product line.