Room 5 · Transformers · 2024

Reasoning Models

Reasoning Models (OpenAI o1 · DeepSeek R1, 2024): More compute spent on intermediate steps before answering.

In September 2024 OpenAI presented o1, a model trained to generate a long chain of intermediate steps before answering. In early 2025, DeepSeek released R1, an open-weights reasoning model. These models scored much higher on mathematics, programming and science tests. The story is still being written.

In 2022 researchers noticed something simple. If you ask a language model to show its work, writing intermediate steps before the final answer, it gets more math and logic problems right. A Google paper led by Jason Wei called this chain-of-thought prompting, and another team found that just adding "Let's think step by step" to a question improved results.

In September 2024 OpenAI presented o1, a model trained to do this on its own. Using reinforcement learning, it learned to produce a long internal chain of reasoning before answering: trying approaches, checking them and correcting mistakes. OpenAI reported that performance kept improving both with more training and with more time spent thinking at the moment of answering. That second point was new. Instead of only making models bigger, you could also let them work longer before answering, spending more computation per question.

On hard math competition problems and scientific questions, these reasoning models scored far higher than earlier chat models. They are also slower and more expensive per answer, and their visible reasoning does not always reflect how they actually reached a conclusion, an open question in interpretability research.

In January 2025 the Chinese company DeepSeek released R1, a reasoning model with open weights and a paper describing how large-scale reinforcement learning had made reasoning behavior emerge. Its reported training cost was far lower than many expected, and its release was followed by a sharp fall in the stock prices of AI chip companies. Within months, reasoning modes became a standard feature in models from Google, Anthropic and others.

This is the last piece of the room because the story is still being written. New models may be in our news by the time you read this.

Ask your AI: does it reason step by step before answering? Ask it how sure we can be that its written reasoning is what it really did.

Did you know? The instruction “let's think step by step” became popular in 2022 as a way to ask a model for intermediate steps; the technique was called “chain of thought”.

Source: openai.com/index/learning-to-reason-with-llms/

See it in the museum →

← ChatGPT

All of Room 5: Transformers

Cite this: Nazareno, I. (2026). Reasoning Models (2024). New Mind Museum. Retrieved November 17, 2026, from https://newmindmuseum.com/museum/reasoning-models-2024/