Pricing
What a Custom GPT or RAG Assistant Costs
The build cost of an assistant is mostly document preparation, and briefs almost never account for that.
The model is a commodity. The interface is a small piece of work. What takes the time is getting your material into a state where the right passage comes back for a real question: cleaning it, splitting it sensibly, handling the tables and diagrams that naive processing destroys, and testing retrieval against questions the people who do the job actually ask.
A quote that treats document preparation as a footnote is a quote that will move.
What the cost is actually made of
Roughly, and it surprises people every time.
Document preparation and retrieval tuning is the largest share. It scales with how messy and varied the material is, not with how much of it there is. A thousand consistent documents are easier than fifty inconsistent ones.
The evaluation set is a real line item and worth protecting. Real questions with known answers, written with your team. Without it there is no way to tell whether a change improved the system or quietly broke part of it.
The assistant itself is comparatively small: retrieval, answering, citation and refusal behaviour.
Instrumentation is small and non-negotiable: metering, caps, breaker, kill switch, logging and feedback capture.
Running cost, and keeping it down
Three levers, and two of them are engineering decisions rather than usage.
How much context each question sends. Retrieving twenty passages where three would do multiplies the cost of every question with no measurable gain. Tuning this is one of the highest-return optimisations available.
Which model answers. On retrieval work a cheaper model frequently scores the same against the evaluation set. Starting expensive and never measuring is a common and avoidable error.
Volume. The only one outside your control, and the one worth forecasting before launch rather than after.
Every deployment logs cost per call from day one, so this is a number you watch rather than a number you discover.
Where a smaller build is the right answer
If the requirement is one person answering questions from a handful of documents occasionally, a general assistant with the files attached may be enough, and I will say so rather than build something.
The case for a proper build arrives when several people need it, when the material is too large to attach, when answers must be traceable to a source, or when the material cannot leave your environment. Those are the conditions that justify the preparation work, and without at least one of them the economics do not hold.
Related
Talk about your situation
The first conversation is short and mostly questions. Get in touch and tell me what you are trying to fix.
Frequently Asked Questions (FAQs)
How much does a custom GPT cost to build?
It is driven by the state of your documents rather than by the model or the interface. Consistent, current material with a clear owner makes for a short build. Material scattered across formats and versions makes for a long one, and that preparation work is worth doing regardless of whether an assistant is built on top.
What are the ongoing costs?
Model usage, driven by question volume, how much context each question sends and which model answers. The second and third are engineering decisions that are tuned during the build and measured against the evaluation set. Costs are metered per call and capped daily, so the figure is visible from the first week.
Is it cheaper to use an off-the-shelf product?
Frequently, yes, and it is worth checking first. Off-the-shelf wins when your requirement resembles everyone else's. A build wins when the material cannot leave your environment, when it must integrate with something specific, or when per-seat pricing exceeds the cost of owning the thing at your headcount.
What happens if it does not work well enough?
The evaluation set is what makes that a measurable question rather than an argument. If retrieval accuracy or answer support falls below the agreed threshold, the diagnosis is in the numbers: usually document preparation rather than the model. That is why the evaluation set is built before launch instead of after complaints.