# Stop overpaying by 50 percent: a cost-quality decision tree for choosing the right AI model tier Source: https://getstartedwithai.net/tutorials/model-selection-framework Site: Get Started With AI Tagline: Build AI automations that actually run your business. Audience: solo professionals, freelancers, consultants and AI implementers in training Contact: hello@getstartedwithai.co.uk This Markdown companion is generated from the prerendered page for AI crawlers. # Stop overpaying by 50 percent: a cost-quality decision tree for choosing the right AI model tier Most teams overpay for AI by 40 to 60 percent. They route every request through a flagship model out of habit, then wonder why the API bill climbs. The fix is not a spreadsheet. It is a two-minute decision at job start: does this task actually need the most capable model available, or will a cheaper, faster tier do the job just as well? This tutorial builds that decision tree. You finish with a working framework you can use yourself and hand to a client ops team on the same day. ## Prep before you build (5 minutes) - Access to at least one AI provider with both a flagship and a lightweight model. Most major providers offer both tiers. - An API key or team account with usage tracking switched on. You need to see real spend per model, not just total spend. - A short list of recurring tasks your team routes through AI today. Five to ten is enough. Pull from your workflow audit if you have one. ## The two-tier model Every provider now ships at least two functional tiers. The exact names and pricing shift, so the principle matters more than the label. - Tier 1 is the cheap, fast tier. These models are 10 to 20 times cheaper per token than their flagship siblings and respond in under two seconds. They handle classification, tag assignment, routing logic, extraction, summarisation of short documents, drafting routine emails, filling in templates and answering yes or no questions. If your output has a clear right answer and a short input, this tier usually gets there. - Tier 2 is the flagship or reasoning tier. These models cost more, take longer, and genuinely outperform the cheap tier on tasks that require multi-step reasoning, long document analysis, careful writing judgment and ambiguous briefs. The extra spend is justified when the cost of a wrong output is high. - Most real workflows need both. The skill is knowing which job goes where. ## Step 1: run the task type test For each recurring AI task, answer four questions in order. Stop at the first yes. 1. Is the input under 2,000 words and the expected output a single, structured answer? If yes, use Tier 1. 2. Does the task involve classification, tag assignment, routing logic or extracting a named field from a document? If yes, use Tier 1. 3. Does the output require reading more than 10 pages of source material, or holding multiple threads together at once? If yes, use Tier 2. 4. Will a wrong output cost more than the price difference between tiers to fix? If yes, use Tier 2. 5. If you reach the bottom without a yes, Tier 1 is probably fine. Test it. Upgrade only when the output fails. ## Step 2: score your recurring tasks Take the five to ten tasks from your prep list. Give each task one point for each of three criteria. - Short input (under 2,000 words) and predictable output: 1 point. - Low cost of error (easy to spot and fix in under 5 minutes): 1 point. - No multi-step reasoning required: 1 point. - A total of 3 points means Tier 1. A total of 2 points means Tier 1 with a spot-check built in. A total of 1 or 0 means Tier 2. - Run this scoring in a spreadsheet. Add a column for current spend per task. The gap between current spend and the tier-appropriate spend is your client pitch. ## Step 3: build the routing rule into your workflow Knowing which tier to use is not enough. You need the routing decision to happen automatically, not every time someone runs a task. 1. For API-connected workflows, add a model parameter at the workflow level, not the prompt level. Every prompt that scores Tier 1 points to the lightweight model slug. Every prompt that scores Tier 2 points to the flagship. One setting, not a manual choice each run. 2. For chat-based team use, create two named shortcuts or saved prompts: one that opens on the cheap tier, one that opens on the flagship. Label them by task type, not by model name. Staff do not need to know the model name; they need to know which shortcut to pick. 3. For client deployments, document the routing rule in the handover pack. The Bus Test question here is: can a replacement team member run this workflow without knowing which model tier to pick? If the answer is no, the routing is not embedded yet. ## Step 4: pressure-test the cheap tier before locking it in Before committing a task to Tier 1, run a side-by-side test. Same prompt, both tiers, 10 representative inputs. Score the outputs against your acceptance criteria, not against each other. - If Tier 1 passes on 9 of 10 inputs, route it to Tier 1 and add a monthly spot-check to your governance cadence. - If it fails on 3 or more, route to Tier 2 and retest in 6 months. Models improve faster than the year before. - One common mistake: teams test with easy inputs and assume the cheap tier handles edge cases equally well. Build your test set from the worst 20 percent of real inputs your team has seen, not from clean examples. ## Step 5: calculate the saving and report it The client does not care about model tiers. They care about the bill and the quality. 1. Take the volume of Tier 2 tasks you are moving to Tier 1 per month, multiply by the token count per task, multiply by the price difference between tiers. 2. A team running 50,000 Tier 2 tokens a day that shifts 60 percent to Tier 1 typically cuts the model spend by 50 to 70 percent with no measurable quality drop on the routed tasks. 3. Report this as cost per output, not cost per token. Clients understand a summary costs 3 pence to produce better than a token-rate quote. ## What you ship A scored task list with model tier assigned to each item. A routing rule embedded in your workflow or team tooling. A before-and-after spend calculation that quantifies the saving. A handover-ready one-pager the client team can run without you. ## What to build next Pair this with the prompt chaining tutorial to handle tasks that start on Tier 1 but escalate to Tier 2 mid-run. The SOP capture build is the right next step if you want to document the routing rule as a client-facing process before you hand it over. ## Related builds More tutorials in Become an Implementer. Each one runs in 60 to 90 minutes and ships a finished asset. ### Land your first paid AI engagement in 90 days: the 2026 consultant playbook 30 to 60 minutes. Implementer. [Open tutorial](/tutorials/how-to-become-an-ai-consultant-2026) ### Ship five repeatable client engagements: the AI automation agency playbook 30 to 60 minutes. Implementer. [Open tutorial](/tutorials/ai-automation-agency-practitioner-playbook) ### Win client work with a free first call: the AI readiness assessment template 30 to 60 minutes. Implementer. [Open tutorial](/tutorials/ai-readiness-assessment-free-template) ### Set rates that pay back in week one: pricing AI implementation work 30 to 60 minutes. Implementer. [Open tutorial](/tutorials/pricing-ai-implementation-work) ## Questions people ask before they start ### How do I know if a model is genuinely Tier 1 or Tier 2? Look at the provider published pricing page, not the marketing names. If one model costs 10 times or more per million tokens than another from the same provider, that gap tells you everything. When pricing is unclear, run the Step 1 task type test and let output quality make the decision for you. ### What if the cheap tier keeps failing on one specific task? Move that task to Tier 2 and do not revisit it for 3 months. Better prompting sometimes closes the gap on borderline tasks, but if 3 of 10 test runs fail at the start, you will spend more on prompt engineering than the model cost saving is worth. ### Can I use this framework with a single-model subscription rather than an API? Yes, but the saving is smaller. Most subscription tiers let you switch between models inside the same interface. The routing rule still applies; you just do it by switching a dropdown rather than a parameter. Document the rule in a team shortcut so the decision is not re-made every time. ### How much should I charge a client to run this model audit? A half-day model routing audit with a scored task list and routing rules embedded in their workflow is worth 600 to 1,500 pounds. The deliverable is a live configuration plus a one-pager, not a slide deck. Anchor the fee on the monthly saving you calculate in Step 5: a client saving 800 pounds a month on model spend will pay 1,200 pounds once for the work. ### Does this framework still apply when providers release new models every few months? Yes. The tier principle is stable even when the specific model names change. Review your routing rules quarterly, test any newly released cheap-tier model against your task set, and update the configuration if the new model passes. The scoring worksheet in Step 2 is the living document; the specific slugs in the API call are just the implementation detail.