Choosing a Model Tier: Fast vs Reasoning (Prompt Guide)
Cleanline runs 22 dry-cleaning stores and its head office routes about 340 items a day through one assistant: nightly store reports in the morning and customer complaint tickets all afternoon. When every item went to the same heavy model, the morning digest arrived after the opening shift had already started and the monthly bill had doubled. The queue was not the problem. Sending a two-line stock count and a 900-word complaint thread down the same pipe was.
Model tier is a routing decision, and it belongs in the prompt, not in a guess. The table below maps each task in the queue to a tier, then a copy-paste prompt follows for each tier, then the cost and latency trade-offs. Where the output has to be machine-readable, keep the schema rules from our structured JSON output prompt beside this one.
Task-to-Tier Decision Table
| Task in the queue | Model tier | Why it fits | What it costs you |
|---|---|---|---|
| Turn 22 nightly store reports into one morning digest | Fast tier | Short inputs, output read in under a minute | Misses cross-store patterns hiding between lines |
| Explain why one store’s repeat complaints doubled | Reasoning tier | Must correlate dates, staff shifts, and ticket text | Slower and pricier on every ticket it touches |
| Sort a complaint into one of five categories | Fast tier | Fixed labels, high volume, low ambiguity | Misreads sarcasm and mixed-language notes |
| Draft a reply about a disputed refund | Reasoning tier | Weighs policy against the customer’s history | Overkill for a simple damaged-garment claim |
| Pull totals from 40 receipts into one row | Fast tier | Structured field extraction with a set shape | Invents a column if the output format is loose |
Ask two questions before you route. Does the task need an answer that already exists in the input, or does it need a judgement that is not written down anywhere? And is the output consumed in seconds by a person, or compared over time by a system? Lookup-plus-speed goes fast. Judgement-plus-scrutiny goes reasoning.
Deciding by Task, Not by Habit
Most teams pick one tier and keep it, which is why Cleanline’s bill doubled without anyone deciding to spend more. The habit is understandable: one model means one set of prompt quirks to learn. The cost is that the cheap, high-volume work subsidises the expensive, rare work in the wrong direction, and the slow tier slows down the tasks that never needed it.
Route by the shape of the task. A nightly stock report is a fixed form with 22 rows: the store id, the item count, the till total. There is no reasoning in it, only extraction, so a fast tier returns it in time for the shift briefing. A doubling of repeat complaints at the Riverside branch is a different animal: the cause is buried in ticket text from three weeks and needs the model to hold several facts at once, which is exactly what a reasoning tier is for.
Watch the input size as well as the difficulty. A morning digest built from 22 short reports is a modest prompt, while a complaint investigation that pulls a month of ticket history can run long. When a single prompt grows past what one pass can hold comfortably, split it, and use the budgeting rules in our context window budget guide to keep the important part of the input near the instruction.
Sort the queue into three buckets and the routing gets easy. Bucket one is extraction: reports, receipt totals, fixed-label sorts. Bucket two is judgement: causes, comparisons, drafts written against a policy. Bucket three is everything that is neither, which usually means the task is badly framed and should be rewritten before it is routed at all. Most “the model got it wrong” complaints at Cleanline traced back to a bucket-three task that nobody had defined, sent to whichever tier happened to be open.
Prompts That Name the Tier
Fast tier for classification. Fixed labels, strict shape, no explanation.
You are sorting daily operations text for a dry-cleaning chain. For each customer complaint, return exactly one category from this list: DAMAGED, LATE, LOST, BILLING, STAFF. Output one JSON object per line with the fields ticket_id, category, and confidence from 0 to 1. Do not explain your choice. Do not add fields. If the text is too short to decide, use category UNKNOWN and confidence 0.Reasoning tier for a diagnosis. Name the reasoning you want back, not just the answer.
Act as an operations analyst for a dry-cleaning chain. Input: four weeks of complaint tickets for the Riverside store, tagged by date, shift, and desk. Question: why did repeat complaints rise this month? Return: (1) the single most likely cause with the ticket numbers that support it, (2) two alternative causes you ruled out and the evidence that ruled them out, (3) the one data point that would confirm or overturn your conclusion. Show your reasoning in a short numbered list before the conclusion.Router prompt. Decide the tier before you spend it.
Look at the incoming task below and route it. Output one word first: FAST if it is a lookup, extraction, or fixed-label sort; REASONING if it needs comparison across several facts, a cause, or a judgement. Then output a one-line reason and the estimated input length in words. Task: classify 60 complaint tickets into five categories.Latency and Cost Trade-Offs
The fast tier buys you speed and volume. It is the right default for anything a person glances at and moves past, and it is why the morning digest reaches the shift briefing on time. Its cost is depth: it will not notice that the same desk name appears in every late-complaint this month.
The reasoning tier buys you depth and pays in seconds and cents. Used on a genuine investigation, that trade is excellent, because one correct diagnosis of a complaint spike is worth more than a fast wrong guess repeated 340 times. Used on a stock count, it is pure waste.
The real risk is silent mismatch. A reasoning-tier prompt run on a fast model still returns an answer, just a thinner one, and nothing flags it. If a routing rule is in play, log which tier handled each task so you can review mismatches later. Cleanline’s rule should be readable at a glance: reports and fixed-label sorts go fast, anything that starts with “why” or “compare” goes reasoning.
A Routing Rule You Can Keep
Write the rule down as a short policy, not a preference. Lookup, extraction, and fixed-label sorting use the fast tier. Cause, comparison, judgement, and drafting against policy use the reasoning tier. Anything above a set input length is split before it is routed, regardless of tier.
Review the split monthly. If the reasoning tier is handling more than a quarter of the queue, either the tasks really are hard or the router is too generous. If almost nothing reaches it, check whether complaints are being flattened into categories before anyone looks for a cause. Then re-run one routed week and confirm the digests still arrive before the shift starts.
Keep the router itself boring. A routing prompt should decide a tier, give one reason, and stop. When it starts explaining its own choice at length, it has become a reasoning task in its own right and is costing you exactly what the routing was meant to save. Log the tier on every task, review the log once a month, and move any task that keeps landing in the wrong bucket by rewriting its prompt rather than by turning up the model.
Sources & Notes
These prompts were written for a dry-cleaning chain routing reports and complaint tickets, but the fast-versus-reasoning split applies to any queue that mixes lookups with judgement. Cleanline, its 22 stores, the Riverside branch, and the 340 daily items are an illustrative scenario, not a real company or figure. Fast and reasoning tiers are described generically; check your provider’s current model names, limits, and pricing before you set a routing rule.
