The Machine That Shows Its Work

The newest AI models "think" before they answer. Here is what reasoning models really do, when the extra cost is worth it, and why showing the work does not guarantee it is right.

The work is visible. That doesn't make it right.

A Different Kind of Answer

For years, AI worked one way. You asked a question and it started generating an answer immediately, one word at a time, no visible pause. Fast, fluent, and for hard problems, often confidently wrong.

Then a new category arrived. Reasoning models do something the older ones didn't. Before giving you a final answer, they generate a long internal train of thought, working through the problem step by step, checking their own logic, sometimes catching and correcting a mistake along the way. Only then do they respond. Some tools even show you this process, a scrolling stream of the model talking itself through the problem.

The results on hard tasks, complex math, multi-step logic, careful analysis, can be dramatically better. But "reasoning" is a generous word for what is actually happening, and it is worth being precise about what is and is not going on.

What "Reasoning" Really Means

A reasoning model isn't thinking the way a person does. There's no understanding, no awareness, no inner voice. What it is doing is generating its own intermediate steps as text, and then using those steps as context for the final answer.

Remember two ideas from recent issues. A model predicts the next words based on patterns, and it works with whatever is in its context window. A reasoning model puts those together in a clever way. By writing out the problem step by step, it fills its own context with useful scaffolding, and each step makes the next one more likely to be correct. It is, in effect, giving itself a better prompt before it answers.

This works, and it is worth being clear about why. The model isn't reasoning its way to the truth. It is writing out one reasonable step after another, and those steps tend to guide it closer to a correct answer than answering in one shot would. Helpful. Not the same as knowing.

You Already Have Both

Here is the part that surprises people. You probably have both kinds of model sitting in the same app.

ChatGPT offers Instant, Thinking, and Pro. Instant is the fast, standard model. Thinking and Pro are the reasoning ones.

Claude has a Thinking option you switch on when you want it to work through a problem before answering.

Gemini calls it the thinking level: Standard for everyday questions, Extended for deeper reasoning, and Deep Think on its top paid plan.

The names differ. The idea is the same. One setting for quick answers, one for careful ones.

🚀 MARTY SAYS

"Out here, a navigator who shows every calculation is easier to trust than one who just calls out a heading. That doesn't make the math right. It just makes the wrong turn easy to spot."

Why It Helps, and What It Costs

The upside is real. On problems with many steps, where a single wrong turn early ruins everything, working through the logic out loud catches errors that a fast answer would sail right past. For math, coding, planning, and careful analysis, reasoning models are a meaningful step up.

The costs are just as real. All that internal thinking is generated text, which means more tokens, which means these models are slower and more expensive to run. And they are overkill for most everyday tasks. Summarizing an email or drafting a note doesn't need a model to deliberate for thirty seconds. Using a reasoning model for simple work is like hiring a structural engineer to hang a picture. Capable, but a waste of time and money.

There is one more trap. A model that shows its work looks more trustworthy, and the visible steps can lull you into skipping your own check. But a reasoning model can still hallucinate, and it can produce a flawed chain of steps that arrives, confidently, at a wrong answer. The work being visible isn't the same as the work being correct.

PICK THE RIGHT TOOL

Reaching for the right tool is most of the skill. Use a reasoning model when the problem has multiple steps, when a small error early would wreck the result, or when you need analysis you can follow and audit rather than a quick draft. Use a standard, faster model for the everyday majority: summaries, rewrites, brainstorming, simple questions. The visible chain of thought is a feature worth using, because it lets you spot where the logic breaks, but read it as an aid to your own judgment, not a replacement for it. A model that shows its work still needs you to check the answer.

Safe Harbor: Three Things You Can Do This Week

Run the two-model test. Pick a question with a few moving parts, like planning a five-day trip on a fixed budget with a couple of must-do stops. Ask your fast model, then ask the reasoning one and watch it work through the constraints. Seeing the two answers side by side is the quickest way to learn when the slower one is worth the wait.
Read the chain of thought when you can. If your tool shows its reasoning, actually read it on an important question. Watching where the logic is strong and where it wobbles builds a sharper instinct for when to trust the answer.
Verify the destination, not just the journey. A visible, sensible-looking set of steps can still end in a wrong answer. On anything that matters, check the final result against a source you trust, exactly as you would with any AI output.

Next week: how a batch of "private" AI chats ended up on Google, without anyone being hacked.