The next leap was not a friendlier chat box. It was more thinking time.
In Part 1 the loop was prompt → fluent reply. That is fast, and it is often shallow on hard problems. The reasoning era is what happens when a model is allowed to spend extra computation before it answers — especially on coding, math-ish puzzles, and multi-part analysis.
You feel it as a pause. Then a tighter plan, a better refactor, or a clearer diagnosis of a bug.
Test-time compute, in plain English
Training a model is expensive and mostly done before the model is released. Test-time compute is extra work at your question: internal steps, search over possible approaches, checking a draft against itself.
OpenAI’s reasoning models guide is the durable public explainer. Reasoning models spend internal tokens planning before they speak. Developers can raise or lower that work with a reasoning.effort setting — more thinking for harder problems, less when latency matters. The same idea showed up in the o-series: extra test-time compute, not a friendlier chatbot.
You do not need the paper’s benchmarks. You need the mechanism: the model is no longer blurting the first plausible paragraph. It is spending tokens on a private scratchpad, then giving you the conclusion.
That is still an answer loop. It is a better answer loop.
Why coding and complex reasoning showed it first
Software and contest math punish fluent nonsense. A function that does not compile is a failed test. A proof with a hole is wrong. Those domains made the pause worth paying for.
Chat was already “good enough” for drafts. Reasoning models earned their keep where one-shot generation kept failing the last 10% — the bug that only appears after three files, the constraint that invalidates the first plan.
What reasoning does not do
Better thinking does not book the meeting, call the API, or change the production database. A reasoning model can plan those steps in text. Hands are a different era.
That distinction matters. People hear “it thought for 40 seconds” and assume agency. Agency is tools plus a loop that acts. Reasoning is extra inference on the way to words.
Example
You paste a failing test and a stack trace. A chatbot-era model guesses a fix. A reasoning-era model walks the call stack, considers two causes, discards one, and proposes a patch that matches the assertion. You still need to apply the fix yourself, run the test, and merge or push.
The intern got a whiteboard. They did not get deploy keys.
Conclusion
Reasoning improves thinking quality. It prepares the ground for agents: better plans are useless if nothing can execute them — and dangerous if something executes them blindly.
Next we give the model hands.
Takeaway: The reasoning era is extra compute on hard problems — still answers, not actions.
Sources
Part 1: The Chatbot Era
Part 3: The Agent Era