Classification Workload Cost Comparison

Compare an LLM, an embedding classifier and a classical model on monthly cost at your volume, with the build effort amortised and the crossover volumes computed.

Items to classify per month

items / mo

LLM route — input tokens per itemInstructions, label list and the item itself.

tokens

LLM route — output tokens per itemA label is a few tokens. If you ask for a justification it is not.

tokens

LLM route — input rate, USD per 1M tokens

$ / 1M in

LLM route — output rate, USD per 1M tokens

$ / 1M out

Embedding route — tokens per item

tokens

Embedding route — embedding rate, USD per 1M tokens

$ / 1M

Embedding route — hosting, USD per monthThe index and the small classifier on top of it.

$ / month

Embedding route — one-off build, engineer-hours

hours

Classical route — compute, USD per 1,000 itemsA linear model or gradient-boosted tree costs almost nothing to run. Almost.

$ / 1k

Classical route — hosting, USD per month

$ / month

Classical route — one-off build, engineer-hoursLabelling, training, evaluation and the serving path. This is where the real cost of the cheap route lives.

hours

Loaded cost per engineer-hour, USDSalary plus employer costs, not salary divided by 2080.

$ / hour

Amortise the build over, months

months

Cheapest at 2,000,000 items a month

$232.67

That is the embedding route. The LLM route costs $2,640 a month at this volume. Price is not accuracy — see below.

LLM route — per item: $0.0013
LLM route — per month: $2,640
Embedding route — per item: $0.0000080
Embedding route — fixed per month: $216.67
Embedding route — per month: $232.67
Classical route — per item: $0.000010
Classical route — fixed per month: $746.67
Classical route — per month: $766.67
Embedding beats LLM above: 165,142 items / month
Classical beats LLM above: 569,975 items / month

What this assumes: that accuracy is equal, which it is not. This page compares price only, and a cheap route is only cheap if it is good enough for the job — a classifier that is three points worse can cost far more in mis-routed work than it saves in tokens. The LLM route assumes one call per item with no retries and no batch discount; apply your batch tier to the rates if you have one. Labelling and evaluation effort belongs in the build hours, and amortising a one-off over some months is a bookkeeping convention rather than a statement about cash.

Three cost shapes, not three prices

The routes differ in shape before they differ in size. The LLM route is almost pure marginal cost: nothing to build, nothing to host, and a bill that scales exactly with volume. The classical route is almost pure fixed cost: weeks of labelling and training up front, then compute so cheap it rounds to zero. The embedding route sits between them, with a small marginal cost and a real but modest build.

That is why the crossover volumes matter more than the monthly totals. Below the crossover, the LLM route wins on cost and wins even harder on time-to-first-result, because it works the afternoon you write the prompt. Above it, the fixed-cost route pulls away and keeps pulling — the gap widens with every additional item, forever.

The honest version of the decision

Start with the LLM route regardless of volume, because it tells you whether the task is even solvable and it produces labelled data as a by-product. If the volume then justifies it, use those labels to train the cheap route and keep the LLM as the fallback for the cases the cheap model is unsure about. That cascade usually beats both pure strategies on cost and matches the LLM route on accuracy, and it is the reason the crossover figures above should be read as a prompt to measure rather than as a verdict.

Classification Workload Cost Comparison

Three cost shapes, not three prices

The honest version of the decision

The reasoning behind this

Related tools