- OpenAI's o1-preview achieved 83% on competition mathematics compared to GPT-4's 13% in September 2024.
- Reasoning models use reinforcement learning on verifiable tasks like math problems and code challenges with unambiguous correct answers.
- OpenAI's o-series, Anthropic's Claude 3.7 Sonnet, DeepSeek R1, Google's Gemini 2.5 Pro, and Mistral's Magistral Medium all offer reasoning capabilities.
- Reasoning models can cost ten to twenty times more than GPT-4o per task due to generating more tokens and slower processing.
- These models excel at mathematics, competitive programming, formal logic, and scientific analysis but are overkill for simple Q&A and writing.
- Adaptive reasoning that dynamically allocates compute based on problem difficulty is emerging as the next efficiency frontier.
The September 2024 Leap
When OpenAI released o1-preview in September 2024, it changed how the AI community thought about model capabilities. Here was a model that, on competition mathematics, went from GPTGPTGenerative Pre-trained Transformer — the model architecture and family name behind OpenAI's most famous models, from GPT-2 to GPT-5.Learn more →-4's 13% to 83%. This was not achieved through bigger training data or more parametersParametersThe numerical weights inside a neural network that are learned during training — the 'knowledge' of the model, measured in billions for modern LLMs.Learn more →, but through a new approach: letting the model think before answering.
Reasoning models marked the shift from 'predict the next tokenTokenThe basic unit of text that an LLM processes — roughly corresponding to a word or word-piece. Models read input and produce output in tokens, which is also how API usage is measured and billed.Learn more →' to 'allocate compute intelligently.' By spending tokens on intermediate reasoning steps, these models could tackle problems that exceeded what any generative model had previously managed. The era of test-time computeTest-Time ComputeUsing additional computation during inference to improve model outputs through techniques like extended reasoning, multiple sampling, or search algorithms.Learn more → had begun.
How Reasoning Models Actually Work
Reasoning models are trained using reinforcement learning on verifiable tasks. These include math problems, code challenges, and formal logic puzzles where there is an unambiguous right answer. The model is rewarded for reaching the correct answer and learns over time to develop reasoning strategies that work, rather than being explicitly taught how to reason.
Crucially, the reasoning process is hidden in most implementations. Users see the final answer but not the chain of thoughtChain of ThoughtA prompting technique that encourages an LLM to reason step by step before giving a final answer, dramatically improving performance on complex tasks.Learn more →. The model may have tried several approaches, caught errors in its own reasoning, and backtracked, all before producing its output. This internal deliberation is what enables dramatically better performance.
Which Labs Have Reasoning Models
OpenAI's o-series pioneered the category. On GPQAGPQAA graduate-level, 'Google-proof' multiple-choice benchmark of hard science questions designed to resist easy lookup and test deep reasoning.Learn more → Diamond (PhD-level science), o3 scores 87.7%, approaching human expert performance. Anthropic's Claude 3.7 Sonnet introduced 'extended thinking,' a visible chain-of-thought that users can optionally see. DeepSeek R1 matched o1's performance at a fraction of the cost as an open-weightOpen-WeightA model whose trained weights are publicly available for download, allowing anyone to run, fine-tune, or build on top of it — distinct from fully open-source (which also includes training code and data).Learn more → model.
Google's Gemini 2.5 Pro includes deep thinking capabilities. Mistral's Magistral Medium is a European-developed reasoning modelReasoning ModelA class of LLMs trained specifically to 'think' through problems step by step using extended chain-of-thought reasoning before producing a final answer, excelling at math, coding, and complex logic.Learn more →. The rapid diffusion of this technique means reasoning capabilities are no longer exclusive to any single lab.
When to Use Reasoning Models
Reasoning models excel at tasks with verifiable correct answers: mathematics, competitive programming, formal logic, and scientific analysis. They also shine at multi-step planningPlanningAn agent's decomposition of a goal into an ordered set of steps or subgoals before or while acting to achieve complex objectives systematically.Learn more →, complex debugging, and any task where checking your own work is important. They are overkill for simple Q&A, writing, and conversational tasks.
The cost is real. Reasoning models generate more tokens and are slower. o3 can cost ten to twenty times more than GPT-4o per task for hard problems. The routing pattern emerging in production: try fast general models first, fall back to reasoning models only for tasks that fail a quality check.
The Future of Test-Time Compute
The scaling lawsScaling LawsEmpirical power-law relationships showing how model performance improves predictably with increases in model size, training data, and compute resources.Learn more → of test-time compute suggest reasoning will continue improving rapidly. More thinking tokens produces better answers. AlphaProof, DeepMind's mathematical reasoning system, achieved silver medal performance on the 2024 International Mathematical Olympiad.
Adaptive reasoning is the next frontier. Models that dynamically decide how much to think based on problem difficulty will be much more efficient. Spending 100 tokens thinking about a simple question and 10,000 tokens on a competition math problem is the ideal behavior. This compute-efficient reasoning is already emerging in frontier systems.