AI Reasoning Models Explained — What Makes Them Different

Thinking Longer, on Purpose

A reasoning model is trained and prompted to work through a problem in explicit intermediate steps — often called “chain of thought” — before producing its final answer, rather than generating a response in one continuous pass. A standard model asked “what's 17% of 640, then subtract that from 500” might jump straight to an answer that's sometimes wrong. A reasoning model will typically work the arithmetic out in visible or hidden intermediate steps first, checking its own math along the way, before giving the final number.

This isn't a different underlying architecture bolted onto the model — it's the same kind of transformer-based language model as any chatbot, just trained to generate a lot more intermediate reasoning text (or a compressed internal version of it) before it commits to a final answer, and given extra time and computing budget at the moment you ask a question to actually do that work.

Test-Time Compute: Spending More Effort Per Question

More “thinking time” at the moment you ask a question, not just a bigger model, is what separates a reasoning model from a standard one.

Most of the AI capability gains between roughly 2018 and 2023 came from making models bigger and training them on more data — what researchers call scaling at training time. Reasoning models introduced a second lever: scaling at inference time, meaning the model spends more computation solving your specific question, rather than the improvement coming entirely from a bigger pretrained model. This is often called “test-time compute.”

Practically, that shows up as a visible delay and, in some products, a “thinking” panel showing the model's intermediate reasoning before the final answer appears. OpenAI's o1, released in September 2024, was the first widely used consumer product built around this idea, followed by o3, Claude's extended thinking mode, Google's Gemini with Deep Think, and the open-weight DeepSeek-R1 in January 2025, which showed the approach could be trained relatively cheaply and still be highly competitive.

Under the hood, the training process that produces this behavior typically uses reinforcement learning: the model is rewarded not just for getting a final answer right, but for the quality and correctness of the intermediate steps along the way, on problems (mostly math and code) where each step can be automatically verified. That's part of why reasoning models tend to show their biggest gains on exactly those kinds of verifiable, step-by-step problems, and smaller gains on tasks like open-ended writing where there's no single “correct” intermediate step to reward.

Not Every Task Needs It

“Showing Its Work” Isn't Always Honest Work

A reasoning model's visible chain of thought looks like genuine step-by-step logic, but research from Anthropic and others has found the displayed reasoning doesn't always accurately reflect what actually drove the final answer — a model can produce a plausible-looking derivation that isn't the real causal path to its conclusion, similar to how a person might construct a tidy explanation after the fact for a decision they made for different reasons. This matters for anyone treating the visible reasoning as a reliable audit trail rather than a helpful, mostly-accurate summary.

Reasoning models also aren't immune to hallucination — they can confidently reason their way to a wrong answer with an internally consistent but factually incorrect chain of steps, which can actually be more convincing (and harder to catch) than a standard model's shorter, more obviously uncertain-sounding wrong answer. And longer reasoning isn't a guaranteed fix for every kind of mistake: a model can spend a full minute working through elaborate steps and still misread the original question, in which case all that extra computation just produces a longer, equally wrong answer. Treat a long, detailed chain of reasoning as a sign the model tried hard, not as proof the final answer is correct, especially on anything where you can't independently verify the result yourself.

Where This Fits With Other AI Terms

Reasoning isn't a separate category from the foundation models and AI models covered elsewhere in this series — it's a training and inference technique applied to a foundation model. o3 is a foundation model that's also a reasoning model; GPT-4o is a foundation model that generally isn't run in a heavy reasoning mode by default. Most major AI labs in 2026 ship both a fast, standard mode and a slower, reasoning-heavy mode of the same underlying model family, letting the product (or the person using it) choose the trade-off between speed, cost, and accuracy per question rather than being locked into one or the other.

How Fast This Category Moved

Frequently Asked Questions