Idea Lab · Oct 20 2026 · Fleet winner · Market brief

Open-Source Coding Model Cost Deflation

Does a cheaper open-weight coding model actually save the fleet money once rework is counted?

What it is

The market brief says open-weight coding models under 32B parameters now do about as well as last generation's frontier models, at much lower inference cost. It gives one sentence and no sources.

This calculator turns that into a decision for the fleet: for a month of coding work, compare a frontier API, a hosted open-weight model, and running a small model yourself. It also asks how much worse the cheap model can be before rework eats the savings.

Example prices · replace with current quotes Coding tokens per month (millions) 60 Frontier API, blended $ per million tokens Hosted open-weight API, $ per million tokens Rented GPU server, $ per month Self-hosted throughput (tokens per second) Extra rework rate with the small model 15% Engineer time per reworked task ($) Tasks per million tokens

Option	Token cost	Rework cost	Total / mo	Note

Our Hostinger and Contabo servers are CPU machines. A 32B model on CPU runs at a few tokens per second, so self-hosting means renting a GPU box, not reusing what we have.

Claimed by the proposal, not verified

Sub-32B coding models "matching previous generation frontier performance". No benchmark, model name or date was given.
"A fraction of the inference overhead". No figure, so every price above is an example you should replace.
Public coding benchmarks are short tasks. Long multi-file agent work is where small models usually fall behind.

What to try

Keep the defaults. The frontier API wins: the token savings are smaller than the cost of fixing 15% more tasks.
Lower rework to 2%. Now the hosted open-weight API wins by a wide margin. The whole question is how much rework the small model really causes.
Self-hosting rarely beats a hosted open-weight API at our volume, because the server costs the same whether it is busy or idle.
Drop tokens to 5 million a month. A rented GPU sits idle and becomes the most expensive option.
First step for real: route one low-stakes agent task type (tests, lint fixes) to an open-weight API for two weeks and log rework.