A doctor does not order a full-body MRI to check a sore throat. They match the test to the question. Model selection is the same skill: the most capable model is rarely the right one, and reaching for it on every task quietly burns budget. This module teaches appropriate compute, so you spend power where the work actually needs it.
Frontier models are extraordinary, and they are metered. Buyers, finance teams, and platform owners are all watching token spend. When a rep points the most expensive model at a task a light model would nail in a second, the answer is not better, and the bill is. Matching the model to the task is now a real professional skill, the same way a clinician learns which test answers the question at the lowest cost and risk.
Running a premium model to reformat a list or classify a call type is like sending every patient for a CT scan. It works, it is slow, and someone pays for it.
Knowing which rung a task belongs on is teachable and portable across Claude, ChatGPT, and Gemini. Same ladder shape, different names.
Right compute means faster answers on routine work and full horsepower saved for the deal strategy that actually deserves it.
Appropriate compute means using the smallest model that reliably does the job well. Think of it as a formulary. A nurse practitioner handles routine visits; you escalate to the specialist for the complex case. You do not book the specialist for a flu shot, and you do not send a cardiac patient to the walk-in clinic. Models work the same way: light models for volume and routine, mid models for real reasoning, premium models for the hard, ambiguous, high-stakes work.
How much genuine reasoning does this actually require. If the answer is "almost none, just do the thing," you are on the light rung. If it is "read a lot, weigh tradeoffs, and get it exactly right," you climb. The task tells you the rung, not the ambition.
Three major families lead the frontier: Anthropic (Claude), OpenAI (ChatGPT), and Google (Gemini). They share the same shape. Learn the rungs once and you can read any of them. Tap a family to see its ladder.
Lightest to most capable: Haiku, then Sonnet, then Opus, then Fable.
GPT-5.6 tiers, launched 2026-07-09. Luna, then Terra, then Sol. Names are durable capability tiers, not version numbers. Fast-moving space, verify current tiers before quoting.
The lesson is portable: pick the family your stack and your privacy needs point to, then apply the same light-to-premium logic.
Light rung: Haiku, Luna. Mid rung: Sonnet, Terra. Premium rung: Opus, Fable, Sol. Learn the three rungs once and you can pick correctly on any platform your customer happens to run.
Real go-to-market tasks below. For each one, pick the rung you would run it on: light, mid, or premium. You will get the recommended rung and a one-line reason. There is no penalty for climbing one rung when unsure; the drill rewards not overshooting to premium on routine work.
Here is the reassuring part. You do not have to choose a model on every single prompt. A well-built harness routes each request to the right rung automatically, the way a triage nurse reads the case and sends it to the right level of care before the patient ever picks a specialist.
Before a prompt runs, a cheap, fast model reads it and decides how much reasoning it actually needs, then routes it to the right tier. A one-line reformat goes to a light model. A full deal-strategy build gets escalated to a premium one. You keep typing your prompt normally; the router quietly matches compute to the task in the background.
Even when a router handles it, understanding appropriate compute makes you a better operator. You will phrase tasks more clearly, recognize when an answer felt over- or under-powered, and design workflows that route sensibly. The router is only as good as the judgment built into it.
SPICED Pulse scores calls at volume. Tagging and segmenting thousands of call moments is light-rung work; the deep coaching synthesis and the executive narrative are premium-rung. Routing each stage to the right tier is what keeps the whole engine fast and affordable at scale.
Run these in your own AI tool. Each is tagged with the rung it belongs on, so you feel the difference between routine and heavy GTM work. Copy, paste, and swap in your real deal.
Read the pasted call transcript and label it as one of: discovery, demo, negotiation, renewal, or check-in. Return only the single label and one short sentence of justification. Do not summarize the call. [paste transcript]
Write a short follow-up email to the buyer after our discovery call. Reference the Pain and Impact they described, confirm the next step, and keep it under 120 words in a warm, direct tone. Give me two subject-line options. Context: [paste your SPICED notes]
Summarize this call into SPICED: Situation, Pain, Impact, Critical Event, Decision. For each element, quote the buyer's own words as evidence, then flag any element that is missing or weak. End with the single most important gap to close on the next call. [paste transcript]
Act as my deal strategist. Using the account research, the three call transcripts, and the pipeline notes below, build a multi-threaded strategy for this opportunity: map the buying group and their motivations, quantify the Impact against a metric the buyer owns, pressure-test the Critical Event, name the top three risks, and give me a sequenced 30-day plan with the owner of each move. [paste research, transcripts, pipeline notes]
Write a board-ready narrative for this strategic account. Synthesize everything below into a one-page story: where the relationship stands, the quantified business Impact we are driving, the risks to the renewal, and the recommendation with its rationale. Executive tone, no filler, every claim backed by a specific piece of evidence from the source material. [paste all account context]
Teach the frontier model of choice, and teach reps not to reach for the heaviest model for a lightweight task. That is appropriate compute: the right diagnostic test for the question, not the most expensive one every time.