| Reasoning in AI refers to a class of language models that generate an extended internal thinking trace before producing a final answer, using inference-time compute to work through multi-step problems that a standard LLM cannot solve in a single forward pass. On AIME 2024, GPT-4o scored approximately 12% while o1 scored 74.3%, using the same underlying model class configured to think before answering. These three building blocks created the category, and o1, DeepSeek R1, and Claude 3.7 each implement the approach differently. The Large Language Models subcategory on Universalnest.com covers every dimension of how reasoning models work, compare, and are deployed. |
Reasoning in AI describes a category of language models built to generate an internal thinking trace before committing to a final answer. The approach uses inference-time compute rather than a single forward pass, producing measurably better results on problems requiring multi-step logic, competition mathematics, and complex code analysis. On AIME 2024, the benchmark gap between standard and reasoning configurations of the same model class was 62 percentage points. This article covers what reasoning models are, the three building blocks that made them possible, and how o1, DeepSeek R1, and Claude 3.7 each implement the approach with meaningfully different designs.
What Reasoning in AI Actually Means: The Difference From a Standard LLM
Reasoning in AI describes language models that produce an internal chain-of-thought before committing to a final answer, working through intermediate steps, checking their logic, and revising before responding. Standard LLMs generate output token by token in a single uninterrupted forward pass with no mechanism to pause, reconsider, or retry.
The core distinction is when the model computes. A base LLM produces each token based on the previous ones and has no internal way to recognise when it is uncertain or wrong. A reasoning model generates a reasoning trace first: a structured sequence of intermediate steps it works through before producing output. That trace represents the model’s working memory for that problem.
A practical analogy helps here. System 1 thinking is fast and automatic; System 2 thinking is slow and deliberate. Standard LLMs use System 1 by default. Reasoning models activate System 2 on demand. The difference is not architectural at the transformer level. A reasoning model and a standard LLM are both transformers trained on text. The reasoning model is additionally trained to produce a long internal trace and is given the inference budget to execute it.
The trade-off is direct. Reasoning traces consume tokens, which means higher latency and higher inference cost. That relationship determines when reasoning mode is worth activating.
The Three Building Blocks That Made Reasoning Models Possible
Reasoning models were built from three ideas assembled in sequence. Each building block solved a specific limitation of the one before it. Understanding the sequence is the clearest path to understanding every reasoning model on the market.
| Building Block | Source | What It Enabled | Key Limitation |
| Chain-of-Thought prompting | Wei et al., Google, January 2022 | “Think step by step” improved multi-step accuracy, effective only above ~100B parameters and only as a prompting trick, not baked into weights | Model had to be told to think; could not choose to |
| Test-Time Compute Scaling | Snell et al., 2024 | More compute at inference = higher accuracy; a smaller model with the right thinking budget can outperform one up to 14x larger | Required models already capable of producing useful thinking traces |
| Reinforcement Learning at Scale | DeepSeek R1 (2025), o1 (2024) | Models trained to produce reasoning traces automatically through trial-and-error reward signals with no step-by-step supervision needed | Higher inference cost; emergent reasoning behaviours not always predictable |
How o1 Works: RL-Trained Chain-of-Thought at Inference Time
o1 is trained via reinforcement learning to produce a long internal reasoning trace at inference, a design that raised AIME 2024 scores from approximately 12% (GPT-4o) to 74.3% using the same model class rather than a larger training run or a different architecture.
Released September 2024 by OpenAI, o1 does not use an explicit search algorithm. Reinforcement learning trains it to perform implicit search through chain-of-thought: exploring paths, backtracking where needed, and converging on an answer. The thinking trace is hidden from users in the standard interface. Users see only the final answer, not the steps taken to reach it. This differs from Claude 3.7, which makes the trace visible.
o3, the successor, reached 91.6% on AIME 2024, confirming that inference compute scaling within the reasoning paradigm continues to yield improvements without architectural changes.
The AIME and GPQA Diamond benchmarks used to measure this performance gap are explained in the Large Language Models subcategory article covering how AIME and GPQA Diamond benchmarks measure reasoning model performance against base LLMs.
How DeepSeek R1 Works: Pure Reinforcement Learning Without Supervised Fine-Tuning
DeepSeek R1 achieves reasoning capability through GRPO (Group Relative Policy Optimization), a form of reinforcement learning where the model is rewarded purely for correct final answers with no supervised fine-tuning of the reasoning trace itself. The model discovered how to reason through trial and error rather than being taught it step by step.
Released in January 2025 by DeepSeek, a Chinese AI lab, it matched o1’s benchmark performance at under $6 million in training cost. DeepSeek R1 is fully open-weight: downloadable, locally runnable, fine-tuneable, and architecturally inspectable without an API key. It scored 79.8% on AIME 2024, 97.3% on MATH-500, and 71.5% on GPQA Diamond.
The DeepSeek-AI technical report (January 2025) describes what the team called an “Aha moment”: reflection behaviours, backtracking, and exploration of alternative approaches emerged spontaneously as emergent properties of RL training, not as explicitly programmed steps. The model was not told to reflect or reconsider. It learned to do so because doing so improved its final answers.
GRPO differs from RLHF (Reinforcement Learning from Human Feedback) in one key way: GRPO rewards final answer correctness while RLHF rewards human preference labels. DeepSeek R1 required neither reasoning labels nor preference data for its capability.
How Claude 3.7 Works: One Model With Two Cognitive Modes
Claude 3.7 Sonnet is the first hybrid reasoning model from a major AI lab: a single neural network trained to support both standard fast responses and extended step-by-step thinking within the same architecture, with users and developers able to toggle between modes and set a thinking budget measured in tokens.
Released February 25, 2025, by Anthropic, Claude 3.7 differs from o1 and DeepSeek R1 in one key way: both are dedicated reasoning models that always generate a thinking trace. Claude 3.7 switches modes, so standard tasks pay no reasoning overhead. Extended thinking mode makes the trace visible to users, which o1 does not.
The performance data from extended thinking is specific. AIME 2024 with parallel extended thinking: 80.0%. GPQA Diamond with parallel extended thinking: 84.8%, which outperformed o3-mini at 79.7%. MATH benchmark: 78.2% in standard mode rising to 91.4% with extended thinking, a 13.2-point gain on competition-level mathematics. SWE-bench Verified: 62.1%.
The thinking budget is denominated in tokens. How context window size determines how many thinking tokens a reasoning model can generate per request is covered in the Large Language Models subcategory guide to context windows.
How o1, DeepSeek R1, and Claude 3.7 Compare at a Glance
| o1 | DeepSeek R1 | Claude 3.7 Sonnet | |
| Released | September 2024 | January 2025 | February 2025 |
| Training approach | RL with thinking trace | GRPO pure RL, no supervised fine-tuning | Hybrid: standard + extended thinking mode |
| AIME 2024 | 74.3% | 79.8% | 80.0% (extended thinking) |
| GPQA Diamond | 78.0% | 71.5% | 84.8% (extended thinking) |
| Trace visible to user | No | No | Yes |
| Open-weight | No | Yes | No |
| Training cost | Undisclosed | Under $6 million USD | Undisclosed |
When Reasoning Mode Adds Value and When It Adds Cost Without Benefit
Reasoning mode adds measurable accuracy on multi-step problems: competition mathematics, complex code debugging, and scientific analysis. On tasks where a standard LLM already performs at or near the accuracy ceiling, extended thinking adds cost and latency without proportionate benefit.
Thinking tokens cost the same as output tokens. A 10,000-token reasoning trace costs the same to generate as a 10,000-token document. Claude 3.7’s GSM8K result shows this clearly. Standard mode: 95.8%. With extended thinking: 98.1%. That 2.3-point gain rarely justifies the added token cost for grade-school-level mathematics in production deployments. On the MATH benchmark, by contrast, the same model rises from 78.2% to 91.4%, a 13.2-point gain that clearly justifies the reasoning overhead.
Claude 3.7’s hybrid design addresses this directly by letting developers route simple requests through standard mode and complex requests through extended thinking.
| Task Type | Reasoning Mode Gain | Worth Activating |
| Competition mathematics (MATH) | +13.2 points | Yes: significant gain |
| Grade-school math (GSM8K) | +2.3 points | No: cost exceeds gain |
| Multi-step code debugging | Measurable | Yes: for complex cases |
| Scientific reasoning (GPQA Diamond) | Significant | Yes |
| Text summarisation | Negligible | No |
| Simple factual Q&A | Negligible | No |
Reasoning models represent a structural shift in how AI capability scales. Training-time compute was the dominant lever from 2018 to 2023. Test-time compute scaling adds a second axis: spending compute when answering, not only when training. That shift connects reasoning models to broader questions about where AI capability scaling goes next, including scientific discovery and complex multi-step planning.
AI Reasoning Models in the Large Language Models Subcategory
The o1, DeepSeek R1, and Claude 3.7 reasoning models sit within the complete Large Language Models coverage on Universalnest.com, where every article in this subcategory builds topical authority from the foundation up.
How AIME and GPQA Diamond benchmarks measure reasoning model performance against base LLMs explains the evaluation framework used to compare every model in this article.
How context window size determines how many thinking tokens a reasoning model can generate per request connects the memory constraint directly to reasoning capacity.
LLMs and reasoning models covered across AI News and Tech on Universalnest.com maps the complete landscape across all 12 subcategories in the AI News and Tech category.
What o1, DeepSeek R1, and Claude 3.7 Confirm About Reasoning in AI
Reasoning in AI is not a separate model category at the architectural level. It is the same transformer configured to spend compute at inference rather than only at training. o1 established the performance ceiling of dedicated reasoning mode on key benchmarks. DeepSeek R1 proved the approach is reproducible at low cost as a fully open-weight model. Claude 3.7 added a third path: a single model offering both fast standard responses and deliberate extended thinking within the same architecture.
Universalnest.com covers the complete Large Language Models subcategory, from how these models are trained and benchmarked to how they reason and are selected for production use in 2026.
Frequently Asked Questions: AI Reasoning Models
What is a reasoning model in AI?
A reasoning model generates an extended internal chain-of-thought before its final answer, using inference-time compute to work through multi-step problems. Standard LLMs generate tokens in a single forward pass with no reconsideration mechanism. The distinction is when the model computes, not its fundamental transformer architecture.
How is o1 different from standard LLMs?
o1 is trained via reinforcement learning to produce a long internal reasoning trace before answering. On AIME 2024, this raised scores from approximately 12% (GPT-4o) to 74.3% using the same model class. The gain comes from allocating inference-time compute, not from a larger or fundamentally different training run.
What is GRPO and why does DeepSeek R1 use it?
GRPO (Group Relative Policy Optimization) rewards the model purely for correct final answers with no supervised fine-tuning of the reasoning trace. DeepSeek R1 used GRPO without step-by-step supervision, and structured reasoning behaviours including backtracking and reflection emerged spontaneously from the RL training process.
What is Claude 3.7 extended thinking mode?
Extended thinking mode is Claude 3.7’s reasoning configuration, where the model generates a visible step-by-step thinking trace before answering. Developers set a thinking budget in tokens. On the MATH benchmark, this raises accuracy from 78.2% (standard mode) to 91.4%, a 13.2-point gain on competition-level mathematics.
When should I use reasoning mode?
Use reasoning mode for multi-step mathematics, complex code debugging, and scientific reasoning where accuracy matters more than speed. On simpler tasks already near the accuracy ceiling, extended thinking adds cost without proportionate gain. Claude 3.7’s GSM8K score improves only 2.3 points with extended thinking enabled.
Is DeepSeek R1 open-source?
DeepSeek R1 is fully open-weight: downloadable, locally runnable, fine-tuneable, and architecturally inspectable without an API key. Released January 2025, it matched o1’s benchmark performance at under $6 million in training cost, making reasoning model capability accessible for self-hosted deployment for the first time.
