Reasoning models vs fast models: which questions deserve “thinking”? (don’t hire a detective to read a clock)
Chinmay Raibagkar · routes lookups to fast models and why-questions to reasoning ones🧠 with a routing cost calculator, not a lecturePublished October 4, 2026
The 60-second version
- Thinking helps on why-questions and multi-table work, and adds cost and wait time to simple lookups for no gain in accuracy.
- Most analytics questions are lookups or comparisons — routing them to a fast model can cut model cost by well over half.
- The fix: route by question type, fix missing context instead of upgrading, and measure corrections per type from your question log.
A marketing team upgrades its AI analyst to the newest reasoning model — the one that "thinks" before it answers. It's the smartest option, so they use it for everything.
A month later, two things are true. The answer to "Why did CAC jump in the second week?" is noticeably better: it checks three tables, spots that a high-CPA campaign scaled on the 9th, and shows its working. And the answer to "What did we spend on Meta yesterday?" now takes 30 seconds instead of 4, costs about ten times as much — and is exactly the same number it always was.
Thinking helps on hard questions and does nothing for easy ones. don't hire a detective to read a clock Most analytics questions are easy. The win isn't choosing the best model; it's sending each question to the right one.
One month, everything on the reasoning model
What "thinking" actually buys you
A reasoning model spends extra tokens working through a problem before it answers: planning steps, trying an approach, checking its own work. That's valuable when the question has several steps. It's wasted when the answer is one well-defined query.
Helps: multi-step questions
"Why did CAC jump?" needs a plan — split by channel, find the mover, check what changed. Reasoning models are much better at holding that plan together.
Helps: catching its own mistakes
Thinking gives the model room to notice a join that doubles revenue or a filter that drops a channel, before it answers.
Doesn't help: missing context
No amount of thinking finds a table the model wasn't told about, or knows your definition of 'active customer'. Reasoning can't replace schema context.
That last point matters. If a reasoning model gets an answer wrong because it was never told what a column means, upgrading the model again won't fix it — schema context will.
Route each question to the right model
Routing means classifying the question first (a quick, cheap step) and then sending it to the model that suits it.
The split is illustrative — measure yours from your question log. In most teams the easy questions dominate.
Simple routing rules
Reporting HierarchyLookup
One metric, one period, a definition that already exists. 'Spend on Meta yesterday.' A fast model answers it as well as any.
Comparison
Two periods or segments side by side. 'This week vs last by channel.' A fast model works; add a quick check that totals match.
Diagnosis
A why-question, several tables, an open path. 'Why did CAC jump in week two?' This is where reasoning earns its cost.
Ambiguous
The question has more than one reading. No model fixes this — ask a clarifying question first, then route.
What does routing save?
Plug in your own numbers. Prices and speeds vary a lot between providers and change often, so use what you actually pay.
Reasoning models use many more tokens per answer
Most questions don't need thinking. Routing cuts the bill by more than half and gives easy questions their speed back.
The model is usually the smaller bill. Each answer also runs a warehouse query, and an unfiltered scan can cost more than any model call. See the real cost of an AI analytics question for both halves.
Measure it, don't guess
Whether reasoning earns its cost on your questions is an empirical question. Log every question with its type, the model used, the wait, the cost, and whether the person accepted the answer. Then compare.
Read it row by row. If lookups on the fast model have the same thumbs-down rate as on the reasoning model, the extra cost buys nothing there. If diagnoses on the reasoning model get corrected far less often, that's where it belongs. For a fixed test set to compare models fairly, use your golden questions.
Model choice habits review once a month →
- Default to fast, escalate to reasoning — start every question on the fast model; send why-questions and multi-table work to the reasoning model.
- Fix context before upgrading — when an answer is wrong because of a missing definition or table, add the context; a bigger model won't find it.
- Track cost and corrections by question type — keep the routing rules honest with the question log, not with benchmark headlines.
Quick gut-check
One question. If you get it, the whole post clicks. 30 seconds, no maths
Your AI answers 'What was our ROAS last week?' wrongly because it used gross revenue instead of net. What's the best fix?
Frequently asked questions
What is a reasoning model?
A reasoning model is a language model that works through a problem step by step — planning, trying approaches, checking itself — before giving its final answer. That extra work usually makes it slower and more expensive per answer, but better at complex, multi-step questions.
Should I use a reasoning model for SQL and analytics?
For multi-step diagnosis — why a metric moved, questions spanning several tables — often yes. For simple lookups with clear definitions, a fast model is usually just as accurate, much faster and much cheaper. Routing questions by type gets you both.
How much more do reasoning models cost?
It varies by provider and changes often, but a reasoning answer commonly uses several times more tokens than a fast one, and can take tens of seconds instead of a few. Measure your own costs from your question log rather than relying on headline prices.
The summary
- Reasoning models help on multi-step questions and catching their own mistakes, not on simple lookups.
- Most analytics questions are lookups or comparisons — a fast model handles them as well.
- Route by question type: lookups and comparisons to the fast model, diagnoses to the reasoning model, ambiguous questions back to the person.
- Missing context is not a reasoning problem; add definitions instead of upgrading.
- Log type, model, cost, wait and corrections to keep routing honest.
Takeaways for your next report
- Thinking helps on why-questions and multi-table work; it adds nothing to a simple lookup.
- Routing easy questions to a fast model can cut model cost by well over half and restore answer speed.
- A wrong answer caused by a missing definition needs context, not a smarter model.
- Ambiguous questions should get a clarifying question, whichever model is on duty.
- Measure cost and correction rates by question type from your own question log.
BigQuery Cost Estimator
Estimate what a BigQuery query will cost before you run it, from validator bytes, with a table-scan simulator and a paste-your-query cost audit.
Chinmay Raibagkar
About author →Founder of DataLens AI. He helps non-technical teams read their ad and database numbers with confidence — which number to trust, what to do next, and what to ignore.