> For the complete documentation index, see [llms.txt](https://docs.thecolliery.org/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.thecolliery.org/benchmarks/coaltipple/readme/routing-savings.md).

# Routing savings

> **What this measures.** CoalTipple's headline move is **delegate-down**: route a big *mechanical* task to a cheap tier instead of grinding it on the expensive main. This benchmark runs the **same** task two ways and prices it:
>
> * **SELF**—the expensive main does it itself (measured: an Opus worker).
> * **DELEGATE**—the main hands it to a cheap new worker + verifies the result (measured: a Haiku worker + an Opus merge-check).
>
> **The honest scope.** CoalTipple is advise-only—it cannot meter tokens itself—so these are **controlled-run** measurements (`subagent_tokens` reported per spawn) × the **published per-token rates**, dated. The saving is real **for big mechanical work**; it is *not* a blanket "X% off everything" (see the floor + the hard gate below).

## Inputs (measured + source-grounded, 2026-06-19)

**Task (identical both arms):** scaffold a TypeScript REST API—3 resources × 4 CRUD Express handlers = **12 handlers**, each typed + validated + try/catch + JSDoc. Pure mechanical bulk.

**Measured token totals** (`subagent_tokens`, this machine, 2026-06-19):

| arm      | model     | tokens                                            |
| -------- | --------- | ------------------------------------------------- |
| SELF     | Opus 4.8  | **71,150**                                        |
| DELEGATE | Haiku 4.5 | **51,490** (also terser for the same deliverable) |

**Published rates** ([platform.claude.com pricing](https://platform.claude.com/docs/en/docs/about-claude/pricing), fetched 2026-06-19—⚠️ version-sensitive, re-fetch before re-citing):

| model     | input $/MTok | output $/MTok |
| --------- | ------------ | ------------- |
| Opus 4.8  | $5           | $25           |
| Haiku 4.5 | $1           | $5            |

→ **Opus is exactly 5× Haiku per token** (both input and output). Routing the bulk down to Haiku pays **1/5** the rate.

## The cost

Code-gen is output-dominated, so this prices the measured totals at each model's **output** rate (the dominant term; the exact input/output split is not instrumented—CT is advise-only—so treat these as representative, not to-the-cent):

* **SELF** = 71,150 tok × $25/MTok ≈ **$1.78**
* **DELEGATE** = 51,490 tok × $5/MTok (Haiku worker) + \~8,000 tok × $25/MTok (Opus merge-check) ≈ $0.26 + $0.20 ≈ **$0.46**

**≈ 74% cheaper** to delegate this task down—driven by the 5× tier ratio, helped by Haiku's terser output. (Counting the real SELF as *inline on the main*—no spawn overhead—nudges it slightly cheaper, \~70%; either way the tier ratio dominates.)

## The honest caveats (why this is not "always cheaper")

1. **The floor / crossover.** A spawn carries a fixed overhead (\~35–58k tokens, measured in prior dogfood). For a **small** task—say a one-line helper—that overhead **exceeds** the saving, so delegating *loses*. This is exactly why CoalTipple has `delegateMinLines`: below the floor it **keeps the task inline** (no loss), above it delegates (the saving above). The benchmark's big task clears the floor; a small one would not.
2. **The sensitive hard gate.** Security / crypto / auth / payments / migrations are **never** delegated down, whatever the saving—a cheaper tier's plausible-but-wrong output (a timing leak, a bad migration) passes tests and ships the bug. CT escalates these UP or keeps them on a capable main. Savings never override correctness on error-not-allowed work.
3. **Voice.** A user-facing deliverable (a translation, a write-up) is not delegated to a cheaper model even when bulky—reviewing it to protect voice costs about as much as redoing it.

## Honest headline

> *On a big mechanical task (a 12-handler API scaffold), routing the work down from Opus to Haiku cut cost \~**70–75%** (Opus is 5× Haiku/token; Haiku also ran terser, 51k vs 71k tokens). The saving holds only **above** the delegate floor—a small task costs more to hand off than to do inline, which is why CoalTipple keeps small/sensitive work on the main. Measured + priced 2026-06-19; rates are version-sensitive.*


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://docs.thecolliery.org/benchmarks/coaltipple/readme/routing-savings.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.