When to use a reasoning model, and when a fast one wins

A reasoning model spends extra effort at the moment it answers. Instead of replying in one pass, it generates intermediate thinking, planning, checking alternatives, and working through steps, before it gives you the visible response. The industry calls this test-time compute: rather than buying accuracy by making the model bigger in training, you buy it by letting the model think longer when it runs. By mid-2026 every major provider offers this with a single knob for how hard the model thinks. OpenAI's GPT-5 family has a fast variant and a thinking variant with a router between them, Google's Gemini exposes a thinking level, and Anthropic's Claude offers extended thinking you can dial up or down. That convergence is exactly why choosing between reasoning models and fast ones is now a practical skill rather than a model-shopping decision.

And it is a skill worth having, because reaching for a reasoning model on the wrong task costs you time and money for no benefit, while skipping it on the right one costs you the answer.

Here's the dividing line that settles most cases: use a reasoning model when a wrong intermediate step would ruin the final answer. If the task is genuinely multi-step and the result is verifiable, where one bad move early poisons everything after it, the extra thinking pays for itself. Logic with dependencies, real math, non-trivial code, analyzing a dataset for patterns and trade-offs, decisions with competing constraints you want pressure-tested before you act. These are the tasks where you'd want to see the work. If instead the task is one-shot, or judged on tone and coverage rather than correctness, the deliberation is wasted and a fast model is the right call. OpenAI's own guidance points the same way: reasoning models shine on complex problem-solving, coding, and multi-step agentic work, and that last one is why the same trait helps an AI agent run a long chain of steps.

Most everyday work is genuinely fast-model work, and using a reasoning model anyway carries a real cost. On hard problems the extra thinking can lift accuracy meaningfully, but it adds several times the latency and the bill. Amazon researchers describe an overthinking problem where reasoning models generate seven to ten times as many tokens as standard models to reach comparable accuracy on simple tasks, and note one model spending seventeen seconds deliberating what one plus one is. Worse, more thinking can actively lower quality on easy work: a study found a reasoning model writing far longer answers to basic math while scoring lower than the plain model. The lesson isn't that reasoning models are bad; it's that effort should match the task. Pay the thinking tax on the minority of work that needs it, and don't pay it on the other eighty percent, which is most of what a team does in a day across every role.

The strongest objection is that this is a non-decision in 2026: the model picks for you. GPT-5's router sends simple asks to the fast model and hard ones to the thinking model automatically, and Gemini and Claude scale their effort to the question. Reaching for a reasoning model by hand can look like a 2024 habit. Concede it plainly, because the routing is real and genuinely useful. But the router is a default, not a guarantee. It optimizes for the average request, and your high-stakes request is not the average; OpenAI even rolled back its first router because it misjudged what people wanted. Every provider still ships a manual override and an effort knob, which they wouldn't if the router were sufficient, and knowing when to override is the skill. The router also can't read intent it can't see: telling the model a task is hard, giving it a clear goal, and asking it to think still changes the outcome. Picking the right tool and steering it well is a human judgment the model can't fully make for you.