# Model Routing, Not Model Selection: The Real AI Cost Lever

> Author: Chris Jon Graf (AI Strategist & CEO)
> Updated: 2026-08-29
> URL: https://ai-outsourcing.ch/insights/model-routing-not-model-selection-the-real-ai-cost-lever

## Summary

Model routing means assigning the cheapest model capable of reliably handling each task—not automatically the strongest one. Simple tasks run on small models like Haiku, most cases on a workhorse model like Sonnet, and only the hardest five percent need premium models like Opus. This typically cuts costs by 60 to 80 percent.

Model routing means deliberately assigning the cheapest model capable of reliably handling each task—rather than reflexively sending every request to the most powerful and most expensive option available. For simple, structured tasks such as classification or data extraction, a small model is almost always sufficient; only the hardest five percent of cases justify a premium model. Applied correctly, this approach cuts AI costs by 60 to 80 percent without any noticeable drop in quality.

## The Expensive Reflex: Always Reaching for the Strongest Model

Many organisations pick a single premium model and route every request through it—from sorting routine emails to analysing complex contracts. The instinct is understandable: the strongest model makes the fewest mistakes, and nobody wants to be blamed for a compromise. But it's the equivalent of taking a sports car to buy groceries. The car handles the trip effortlessly—but a compact car would have done the job just as well, at a fraction of the cost.

**100×** — Price gap between GPT-4o-mini and o3 at OpenAI (input plus output)

The numbers behind this are stark. At Anthropic, Opus 4.7 costs roughly five times as much per token as Haiku 4.5. At OpenAI, the spread is even wider—about a hundredfold between GPT-4o-mini and o3. In agentic workflows that resend the full context at every step, this gap compounds across iterations. Without deliberate routing, organisations end up paying for compute the task never needed.

## Three Tiers of Intelligence—and What They're Actually Good For

In practice, it helps to think in three tiers rather than a single 'best' model:

- Entry tier (Haiku, GPT-4o-mini): classification, tagging, data extraction, simple summarisation—tasks with clear structure and little room for interpretation.
- Workhorse tier (Sonnet, GPT-4o): the default for most business processes—customer communication, research, drafting, mid-complexity code review.
- Premium escalation (Opus, o1/o3): the demanding five percent—multi-step reasoning, high error costs, strategic decisions, complex planning.

The quality gap between tiers is smaller than most assume. On classification and extraction tasks, the smallest model matches the workhorse model (Sonnet) over 95 percent of the time. Even on harder coding benchmarks, Haiku trails Opus by only about 7.6 percentage points—a difference that rarely matters for routine work, while running many times faster and at a fraction of the price.

## What This Actually Costs: A Worked Example

**60–80%** — typical cost reduction from planner-executor routing versus a single-model approach

One proven pattern works like this: a strong model such as Opus handles planning—breaking a complex task into sub-steps. The actual execution of each step is then handled by cheaper models like Sonnet or Haiku, depending on difficulty. This planner-executor pattern typically cuts overall costs by 60 to 80 percent, because the expensive model is invoked once rather than at every single execution step.

## A Decision Matrix for Your Organisation

A simple rule of thumb helps you get started:

1. Is the task structured with a clearly verifiable answer? → Entry tier.
2. Does it require contextual understanding, tone and nuance, but not multi-step reasoning? → Workhorse tier.
3. Does an error carry high cost or risk, or does the task require multi-step planning? → Premium escalation.
4. Is the task new with an unknown error rate? → Start with the stronger model temporarily, then downgrade once real data is available.

This matrix isn't a fixed rulebook—it's a starting point. Models, pricing and capabilities shift every few months, and any routing scheme set once and never revisited will quietly lose efficiency over time. Treating model choice as an ongoing decision, not a one-off setup, is what keeps the cost advantage intact.

> **The Biggest Lever**
>
> Structured routing—by task type, by confidence score, or via a lightweight classifier model—is the single most effective cost lever in any production AI deployment. It outperforms almost any pricing negotiation with a vendor.

## Where Companies Typically Start

The most pragmatic starting point is rarely a full routing system—it's an honest inventory: which tasks currently run through the most expensive available model despite being structured and repeatable? This first sorting exercise usually uncovers most of the savings potential, long before any classifier or routing logic gets built.

It's worth keeping the model-tier question separate from the vendor question. Whether an organisation ultimately standardises on Claude, GPT or Gemini is a different decision from which tier within a model family fits which task. Separating these two questions leads to a more durable routing setup than conflating them.

Swiss companies in particular tend to move more cautiously from strategy to execution than international peers—a pattern explored in [a Swiss AI podcast on the biggest strategic mistakes in the global AI race](https://www.ki-podcast.ch/ki-standort-schweiz-kmu-strategie-und-globales-rennen), which argues that hesitation, not the technology itself, is the real competitive disadvantage.

## The Next Step

Building a working routing system takes more than a spreadsheet of model names: it requires monitoring, fallback logic for uncertain cases, and the discipline to review the assignment regularly. This is exactly where an AI outsourcing approach adds value—an external division that runs this routing for you, continuously optimises it, and carries the cost accountability, without requiring you to build an in-house team from scratch.

## FAQ

### What is model routing in AI applications?

Model routing is the deliberate assignment of different AI models based on task type—simple, structured tasks run on cheap, fast models, while complex or high-risk tasks run on more capable, more expensive models. The goal is the best cost-to-value ratio instead of using one model for everything.

### How large is the price difference between small and large AI models?

At Anthropic, the factor between Haiku and Opus is roughly fivefold per token. At OpenAI, the spread between GPT-4o-mini and o3 is far wider, at around a hundredfold. In multi-step agentic workflows, this difference compounds further.

### Do cheaper models produce noticeably lower quality?

On clearly structured tasks like classification or extraction, small models match the workhorse model (Sonnet) over 95 percent of the time. Even on more demanding tasks, the quality gap is often only a few percentage points—while running significantly faster and cheaper.

### How much can companies actually save through model routing?

In practice, a structured routing system, such as a planner-executor pattern, typically reduces overall costs by 60 to 80 percent compared to using a single premium model throughout.

### Where should a company start with model routing?

The most pragmatic starting point is an inventory: which repeatable, structured tasks currently run unnecessarily through an expensive premium model? These cases can usually be switched to a cheaper model immediately, without a large project.

## Sources

- [Claude Opus vs Sonnet vs Haiku: how to choose | aiarch.dev](https://aiarch.dev/claude-model-selection)
- [Claude Opus vs Sonnet vs Haiku: when to use which | Cadence blog](https://cadence.withremote.ai/blog/claude-opus-sonnet-haiku)
- [Claude Model Routing: The 2026 Cost Lever | PADISO Blog](https://www.padiso.co/blog/claude-model-routing-cost-lever-2026/)
- [Haiku vs Sonnet vs Opus: Which Claude Model? | ClaudeGuide](https://claudeguide.io/claude-haiku-sonnet-opus-which-model)
- [Practical Guide for Model Selection | OpenAI Cookbook](https://developers.openai.com/cookbook/examples/partners/model_selection_guide/model_selection_guide)
- [GPT-4o vs o1 vs o3: Which OpenAI Model to Use When | MasterPrompting](https://masterprompting.net/blog/gpt-4o-vs-o1-vs-o3-which-openai-model-when)
- [GPT Models Comparison 2026 | CV Infotech](https://www.cvinfotech.com/ai-models/gpt-models-comparison/)
- [An analysis of OpenAI models: o1 vs GPT-4o | PromptLayer](https://blog.promptlayer.com/an-analysis-of-openai-models-o1-vs-gpt-4o/)