Three-Tier AI Model Routing System
Route each AI task by cost, speed, intelligence, and reasoning depth
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 99%
The Three-Tier AI Model Routing System classifies models by the kind of work they should receive. Tier one contains inexpensive, fast working models for summarization, extraction, preprocessing, and other high-volume tasks. Tier two contains capable general models used interactively for everyday writing, questions, coding, and conversation. Tier three contains reasoning models that are slower and more expensive but better suited to difficult analysis, planning, and consequential decisions. For every assignment, evaluate complexity, required turnaround, stakes, input size, modality, and budget, then send it to the lowest tier capable of producing a dependable result. This creates an operating portfolio rather than forcing one model to handle everything.
Origin
Extracted from Marketing Against The Grain, where Sully Omar describes the three-tier system used personally and within AutoGrid to navigate a crowded AI-model market.
Core principles
- 01Use the cheapest adequate model
- 02Fast models handle volume and preprocessing
- 03General models support interactive daily work
- 04Reasoning models are reserved for difficult analysis
- 05Model choice should follow the task rather than brand loyalty
How to run it
- 1
Classify the task
Determine whether the assignment is routine processing, interactive knowledge work, or difficult reasoning. Consider both complexity and the cost of an incorrect answer.
Pro tip Classify the task before opening a specific AI application.
Watch out Familiarity with one model can bias routing decisions.
- 2
Use working models for volume
Send simple, repetitive, and context-heavy preprocessing to fast inexpensive models. Examples include summarization, extraction, formatting, and basic classification.
Pro tip Batch similar low-risk tasks to improve cost efficiency.
Watch out Do not expect these models to resolve nuanced strategic trade-offs.
- 3
Use general models for interaction
Choose a capable everyday model for conversations, drafting, routine coding, and iterative refinement. Favor responsiveness when rapid back-and-forth improves the work.
Pro tip Choose among general models based on strengths such as writing quality or integrations.
Watch out A polished conversational response is not necessarily a rigorous analysis.
- 4
Use reasoning models for hard decisions
Reserve slower reasoning models for complex analysis, planning, and problems that benefit from deeper evaluation. Supply comprehensive context and tolerate longer response times.
Pro tip Use premium reasoning only where the expected value exceeds its extra cost.
Watch out Do not treat a reasoning model like a fast chat model.
- 5
Review special capabilities
Adjust the route when native multimodality, long context, web access, enterprise data, or application integrations materially affect performance.
Pro tip Record recurring task-to-model assignments for the team.
Watch out Model rankings become stale as products change.
In the wild
A fast working model summarizes and structures a four-hour transcript. A general model helps an editor discuss themes and draft copy, while a reasoning model evaluates the most difficult strategic claims only if necessary.
→ The editor receives useful output without paying premium-model costs for every token.
A team uses an inexpensive model to clean reporting exports, a conversational model to refine the planning request, and a reasoning model to evaluate competing quarterly initiatives.
→ Each stage uses the model characteristics best suited to its demands.
Common mistakes
Using premium reasoning for everything
Slow expensive models add little value to straightforward transformations. Route simple work to a cheaper tier.
Using a working model for strategy
Fast models may summarize inputs correctly while failing to reason through dependencies, trade-offs, and risks.
Choosing by vendor alone
Different models from the same or competing vendors have distinct strengths. Route according to the task and capability required.
Is it for you?
Best for
It is best for individuals and teams that use several AI models across high-volume, conversational, and analytical tasks.
Not ideal for
It is less useful when access is restricted to one model or when all tasks have nearly identical complexity and risk.
From the transcript
“I call it a three-tiered system to like thinking about models.”
“You have your working models, which are the ones that are really cheap, really fast.”
“Then we have the reasoning models”
From the episode
Which AI Model Should You Use? (Claude vs GPT & O1 Pro Live Prompt Guide)