Reasoning models show their work.

“Should we put this on one of the new thinking models?” It comes up on nearly every Artificial Intelligence (AI) build now. A reasoning model works a problem out before it answers, the way a careful person scribbles on scratch paper instead of blurting the first thing that comes to mind. That extra step is real, and it earns its keep on genuinely hard problems. On easy ones it mostly just runs up the bill.

What does a reasoning model do differently?

A standard model answers in one pass. It predicts the reply straight through, left to right, closer to autocomplete with a diploma than to a deliberate thinker. A reasoning model puts a step in front of that answer. OpenAI says its models use internal reasoning tokens before producing a response¹, which lets them inspect alternatives, recover from ambiguity, and solve harder multi-step tasks¹. Google describes the same move: with a thinking model, Gemini reasons internally before responding². None of this is new. A 2022 paper found that writing out a series of intermediate reasoning steps³ (a chain of thought) significantly improves the ability of large language models to perform complex reasoning³, and that these abilities emerge naturally in sufficiently large language models³.

How does a reasoning model reach an answer?

What's Real?

• •The gains are real, and they cluster on hard, multi-step work. OpenAI says reasoning models are for complex problem solving, coding, scientific reasoning, and multi-step agentic workflows¹; Google points to complex tasks such as coding, advanced mathematics, and data analysis². It's the same effect the 2022 chain-of-thought work measured: room to lay out the steps helps most when the steps are the hard part.
• •The cost is real, and it's counted in tokens you never see. The hidden reasoning is metered. OpenAI notes those tokens are not visible via the API¹ yet are still billed as output tokens¹; Google prices a thinking request as the sum of output tokens and thinking tokens²; Anthropic bills the same way and warns of diminishing returns that depend on the task, and at the cost of increased latency⁴. Priced in tokens, exactly as the token brief laid out⁵, a thinking step can cost several times a plain answer.
• •On easy problems, the extra thinking can backfire. Researchers named it “overthinking,” where excessive computational resources are allocated for simple problems with minimal benefit⁶. OpenAI's own guidance agrees: for latency-critical tasks that do not benefit from any reasoning¹, leave it off. A plain model is often the right call.
• •The visible “reasoning” is not a transcript of the truth. Anthropic found no specific reason why the reported Chain-of-Thought must accurately reflect the true reasoning process⁷; in one study, a substantial majority of answers⁷ were unfaithful. The steps read like an explanation. Treat them as a story the model tells about its answer, not an audit log of how it got there.

The “so what” — for anyone serving tech clients

• •Match the model to the step, and price it that way. Reserve reasoning for the genuinely hard, high-stakes work: the thorny analysis, the tricky bit of code, the plan a person will act on. Run the routine steps on a plain, faster, cheaper model. Because the hidden thinking is billed as output¹ and priced in tokens⁵, quoting “put it all on the thinking model” can quietly multiply a client's bill for output nobody reads.
• •When a client asks for “explainable AI,” be precise about what the reasoning trace is. It looks like an audit trail, and it demos beautifully, but the model's own account of its steps is not guaranteed to be faithful⁷. When correctness has to be provable, the proof is a test you run, the kind the evals brief describes⁸, not the story the model narrates about itself.