# AI Reasoning Models [[Large Language Models (LLMs)|LLMs]] that spend explicit inference-time compute generating an internal chain of intermediate steps before producing the final answer. Where a classic LLM emits the answer in a single forward pass, a reasoning model first emits a private (or visible) "thinking" trace — drafts, branches, self-corrections — then summarises it into the user-facing response. The headline mental model: **trade more tokens at inference for better answers, instead of more parameters at training**. ## How they differ from regular LLMs - **Inference-time compute scaling.** Quality improves with the length of the thinking trace, not just the size of the model — the inference-time analogue of [[AI Scaling Laws|scaling laws]]. - **RL on reasoning traces.** They're typically post-trained with reinforcement learning on verifiable problems (math, code, logic) where correctness can be checked automatically. - **Visible vs hidden thinking.** Some models expose the trace (DeepSeek R1, open Qwen reasoners); others hide or summarise it (OpenAI o-series). - **Latency and cost.** A single answer can burn tens of thousands of "thinking" tokens. Cost-per-correct-answer matters more than cost-per-token. This is the productionised cousin of [[Chain-of-Thought (CoT) prompting]]: instead of relying on the user's prompt to elicit step-by-step thinking, the model is trained to do it natively. The next step researchers are pushing is [[Latent Reasoning]]: thinking in hidden vectors instead of written tokens. It's cheaper and faster, but you can no longer read the trace, which is a real loss for oversight. [[Hybrid Reasoning Agents]] try a middle ground: think in latent space, act through text. ## The convergence — separate model → unified mode The first generation shipped reasoning as a **separate model**: OpenAI o1 / o3 alongside GPT-4o, DeepSeek R1 alongside V3, QwQ alongside Qwen. Users had to pick the right model up front. The second generation collapses this into **a single model with a switchable thinking mode**: - Anthropic [[Claude]] — extended thinking toggle on Opus / Sonnet. - OpenAI GPT-5 / GPT-5.5 — unified, with a `reasoning_effort` parameter. - [[DeepSeek v4]] — folds the R line into V4 with Thinking / Non-Thinking modes. - [[Mistral Small 4]] — `reasoning_effort` from `none` to `high`. The pattern is now industry-default: one base model, dial reasoning effort up or down per request. Operationally it's much simpler — one routing decision, one billing tier, one memory of the conversation. ## Where they earn their keep - Math, formal proofs, and competitive programming. - Multi-step debugging and refactoring across many files. - Strategic / planning tasks where wrong-direction-early compounds. - Tasks with verifiable answers (the regime RL post-training is built for). ## Where they don't - Latency-sensitive interactive UX — the thinking trace destroys time-to-first-token. - Open-ended creative writing — reasoning models often over-correct toward "safe" outputs. - Cheap classification, extraction, summarisation — a non-thinking pass is faster and just as accurate. - Anything where the user pays per-token without per-answer value being measured. Long traces also fail in their own way: they loop, or never close. A third option, besides turning thinking off or lowering the effort, is to post-train the model to think less. [[ThinkingCap]] does that on open-weight models: on [[Gemma 4]] 26B-A4B it cuts thinking tokens by 38.9% on average and looping from 3.6% to 0.8% of answers, with about the same accuracy on BottleCap's benchmarks. Users on Hacker News report that such models lose some quality outside benchmarks, so test on your own tasks. The right default is **non-thinking by default, thinking when verifiability or planning depth justifies the cost**. Pretending every request needs reasoning is a fast path to a bloated [[AI Cost Management|AI bill]]. ## References - ## Related - [[Large Language Models (LLMs)]] - [[Chain-of-Thought (CoT) prompting]] - [[AI Scaling Laws]] - [[AI Cost Management]] - [[AI Inference]] - [[AI Frontier Model]] - [[DeepSeek v4]] - [[Claude]] - [[Mistral Small 4]] - [[Latent Reasoning]] - [[Hybrid Reasoning Agents]] - [[ThinkingCap]]