AI app development cost: what it takes to build and run

An AI app has two costs: building it, and running it every month once people use it. In my own work, a new AI product built from an empty repo to production typically costs $15k-$40k over 8-14 weeks, and adding one AI feature to a product you already run typically costs $4k-$10k over 2-5 weeks. The running cost is mostly the model API bill, and for the same app it can range from tens of dollars a month to thousands, depending on usage and which model answers.

What the build pays for

The model call is the smallest part of an AI product. Most of the build goes into everything around it, the parts that decide whether the product survives real users:

• Retrieval over your own content, so answers come from your documents and data, with each customer's data kept apart.
• Model calls behind an interface, so the vendor can change later without a rewrite.
• Usage metering and limits per customer, so cost per request stays visible and one heavy user cannot run up the bill.
• Accounts, roles and billing.
• A deployed production system, with the failure cases handled: timeouts, refusals, rate limits and empty replies.
• A way to tell whether the answers are right, usually a test set built from real questions and run on every change.

My own multi-tenant AI support platform, Evoriqa, has all of these, and went from first commit to live in about two months.

Typical build costs by kind of work

These are the ranges I publish for my own work. They are typical ranges, not quotes, and scope sets where in each range a build lands.

The second line is where most companies should start. If you already have a product with users, adding one well-built AI feature proves the value for a fraction of a new product's cost.

What pushes a build up its range

A handful of requirements each add real work, and the more of them apply, the higher a build lands:

• Separate customers whose data must stay apart. Multi-tenancy has to be designed into retrieval, storage and permissions from the start.
• Usage limits or billing per customer. Metering every AI action and turning it into credits, caps and invoices.
• Integrations with systems you already run. Each one brings its own authentication, data shapes and failure modes.
• Compliance such as HIPAA or SOC 2. Audit logs, data handling rules and vendor agreements.
• Voice or real-time features. Streaming, latency budgets and a lot more ways to fail.

What it costs to run each month

Once it is live, the main running cost is the model API bill: how many model calls the app makes, times the tokens each carries, times the price per token.

Take 1,000 active users who each ask 30 questions a month of their documents, with each answer sending about 4,000 tokens (instructions, retrieved passages and the question) and returning about 500. That is 30,000 model calls a month:

• On a small model such as GPT-6 Luna: about $20 a month
• On a mid-priced model such as Claude Sonnet 5.5: about $390 a month
• On a top-tier model such as GPT-6 Astra: about $1,950 a month

An agent costs more to run than a chatbot, because it calls the model several times for one task, to plan, act and check, and each call resends the context. At eight calls per task, the same usage costs roughly eight times as much. On top of the API bill come hosting, the database or vector store, embeddings for your documents, and monitoring. The API pricing calculator works out the API bill for exact token counts on every current model.

Ways to spend less

• Start with one feature, not a platform. Prove the value in a product you already run before building a new one.
• Use retrieval before fine-tuning. For facts that change or differ per customer, retrieval is cheaper to build and to keep current. RAG vs fine-tuning sets out when each fits.
• Keep vectors in the database you already run. If you are on PostgreSQL, pgvector often removes the need for a separate service, as pgvector vs Pinecone explains.
• Use the smallest model that passes your tests, and a larger one only for the steps that need it.
• Meter usage from the first customer, so cost per request never becomes a surprise on the invoice.

Estimate your own

The AI app cost estimator puts your scope and expected usage into a build range and a monthly running cost. If you are not sure you are ready to build at all, the AI readiness assessment takes three minutes and shows the gaps to close first.

When you have a product in mind and want it built, here is how I build AI products: from an empty repo to production, with metering, billing and a codebase the next engineer can read.