Reasoning model
A model that burns extra computation working a problem out before it replies. Slower and pricier, much better on the hard stuff.
See it
What it is
A reasoning model is trained to spend extra computation working a problem out before or while it answers, instead of committing to the first plausible continuation. Providers usually meter that work as reasoning tokens: you wait for them and you are typically billed for them. What they contain, how much of it you can see, and how it is priced all differ by provider, and some show you a summary rather than anything raw. Most expose an effort or thinking-budget knob, which is the actual dial between cost and quality.
Reach for one when the answer depends on steps rather than recall: multi-file refactors, tricky debugging, maths, planning, comparing options under constraints, ambiguous specs where the win is noticing the ambiguity. Skip it for extraction, classification, chat, autocomplete and anything a user is watching a spinner for. A cheap fast model plus a good prompt beats a reasoning model on easy work, and the time to first token on a hard reasoning request can be tens of seconds.
Gotcha: the prompting habits that helped older models can hurt these. 'Think step by step' and elaborate chain-of-thought scaffolding duplicate what the model already does, and micromanaging its process fights the training. Give it the goal, the constraints, and what 'correct' looks like, then get out of the way. Second gotcha: the cost surprise is invisible in the transcript, since a 200 word answer can hide 8000 thinking tokens behind it.
Ask AI for it
Add model routing to this app so hard requests go to a reasoning model and everything else does not. Classify each incoming request with the cheap fast model into 'simple' or 'multi-step', route multi-step to the reasoning model with a medium thinking budget, and expose a 'think harder' toggle that raises it to high. While reasoning runs, stream a status line with elapsed time instead of an empty spinner, since time to first token will be long. Log reasoning tokens, output tokens and cost per request separately wherever the provider reports them in the usage payload, so the thinking spend is visible in the dashboard rather than buried in one total.