AI development cost: what you pay for and how to budget

At Ronas IT, prices start at $3,000 for an AI chatbot or another AI feature in your product or internal process, at $5,000 for a new product built around AI, and at $15,000 for a custom machine learning model. Before the build, you can add an optional analysis phase for $1,000 to $2,000. After launch, you also pay usage fees to the model provider and pay for hosting.
This guide is for founders and managers of small and mid-sized companies who need a budget for an AI project. It covers prices by type of work, what moves the price, which approach costs least, what you pay every month after launch, how to keep the cost down, and how to compare quotes from AI development companies.
How much AI development costs by type of project
The price depends first on what you build. An assistant that answers questions is the smallest project. An agent that decides its own next steps in your systems, or an AI feature that becomes part of your product or internal process, costs more. A custom model that predicts something from your data costs the most. The table lists our starting prices and when each line applies.
| What you build | When it fits | Starting price | Timeline |
|---|---|---|---|
| Analysis phase: the AI tasks defined and checked for technical feasibility, a review of the architecture, and a written scope with a budget and a timeline | Optional, before the build, when you have no written scope | $1,000 to $2,000 | At least 1 week |
| Proof of concept as a model check: one task tried on your own examples | Optional, when nobody knows yet whether a model can do the task well enough | Part of the analysis phase | At least 1 week, within the analysis phase |
| Proof of concept with working code | When the check needs working code, not only an answer | From $5,000 as a basic MVP | At least 2 weeks |
| AI feature: a chatbot or assistant on your website, app, or messenger, or another AI feature in your mobile app, web product, or internal process, including a custom AI agent. The price covers the connections to your systems that already have an API | The AI answers questions, qualifies requests, chooses its own steps across your systems, or works as a feature inside your product or process | From $3,000 per AI feature | At least 1 week |
| New product built around AI | There is no product yet, and the AI is its core | From $5,000 as a basic MVP, from $10,000 as an extended MVP | At least 2 weeks as a basic MVP, at least 5 weeks as an extended MVP |
| API development: a new API for one of your systems | Only when a system the AI needs has no API | From $2,500 | At least 2 weeks |
| Machine learning: a custom model trained on your data | The task is to predict something, such as demand or churn, not to write text | From $15,000 | At least 6 weeks |
| Support: monitoring the live system, fixing issues, and updating it | Optional, after launch, as a monthly plan | From $1,000 per month for an app, from $2,000 for a whole system | Ongoing |
Every build price already includes design, project management, and testing. If the AI feature needs new screens, such as a page where a person approves what the AI prepared, their design is part of that feature's price.
A new web product with AI at its core is one line, not a product plus an AI feature: a basic MVP with the AI inside starts at $5,000 for one platform, such as a web app. A mobile app for the same product adds from 50% of that price.
When an AI feature must reach one of your existing systems that has no API yet, one new API and the feature itself together start at $5,500, or at $6,500 with the optional analysis phase.
What drives the cost of an AI project
Six things move the price the most.
- Approach to the model: an API call to a ready model, search over your documents, fine-tuning, or a model of your own. The next section compares what each adds to the cost.
- State of your data: if the knowledge sits in one database or document store with an API, the AI reads it directly. If it is spread across spreadsheets and mailboxes, part of the budget goes into collecting and cleaning it first. We estimate that work together with the scope of the AI feature.
- Systems to reach: every system the AI reads from or writes to is integration work, and a system with no API needs one first.
- Freedom to act: an assistant that only drafts answers for a person costs less than an agent that changes records. Each action the agent can take needs limits and logs, and it may need a screen where a person approves it.
- Cost of an error: a wrong draft for a support specialist costs a minute of editing, while a wrong amount on an invoice costs money. The more an error costs, the more examples the AI must pass before launch, and the longer testing takes.
- Data rules: when your data may not leave your servers, for example under a client contract, the model runs on servers you control. The servers then replace the usage fee as a running cost, as the section on running your own model below shows. You pay the server provider directly, and that bill depends on the model and the load. Setting up new infrastructure is DevOps work estimated by its scope. Running the project on it, from environments to monitoring, is our DevOps and hosting service, from $200 per month.
How AI agents shape the price of a build
How the vendor builds also moves the price. Our engineers and designers work with AI agents in analysis, design, coding, testing, release, and support: with the agents, they write and test code, build screens, review each commit, and watch the health of running projects. Engineers plan how the idea will be implemented and are responsible for whether it works and how well: they set the agents' tasks, make the architecture decisions, and check and accept the result.
Our engineers work with AI agents, and that is reflected in our prices, most of all on standard scope: websites, MVPs, APIs, and AI integrations such as a chatbot connected to your systems. On a complex product, much of the price comes from compliance, from making every component work correctly with the others and with outside systems, and from domain and business logic. Engineers use agents there too, but the decisions and the responsibility stay with people.
Which approach costs the least: an API, RAG, fine-tuning, or your own model?
These four options answer different questions. Instructions, RAG, and fine-tuning decide how the model learns about your business. An API or your own servers decide where the model runs. A custom model is a different kind of project: it predicts a number instead of writing text.
How the model learns about your business: instructions, RAG, or fine-tuning
A ready model with good instructions costs the least to build, and most text tasks should start there. Each next step adds cost and has to earn it on your own test examples.
“Run your own test examples through a ready model before anyone proposes more. Those results are your baseline, and document search or fine-tuning has to beat it by enough to pay for itself.”
Evgeny Leonov, CTO at Ronas IT
| Approach | What it is | What it adds to the cost | When to choose it |
|---|---|---|---|
| Instructions and examples | Each request carries the instructions, a few examples, and the data of the case | The integration, then a usage fee for each request | Most text tasks: drafting, sorting, pulling fields out of documents, answering questions |
| RAG (retrieval-augmented generation) | Before each request, the system finds passages in your documents and adds them to it | A search index of your documents and more text in every request | Answers have to rely on your own documents, such as policies or a product catalog |
| Fine-tuning | A ready model is trained further on your own examples | Collecting and checking examples, training runs, and new runs when you change the model | Instructions and RAG still fail your test examples on format or style |
Our generative AI integration guide explains how each approach fits into a product.
Where the model runs: the provider's API or your own servers
Through an API, you use a model such as Claude, GPT, or Gemini on the provider's servers and pay for each request. With an open model, such as gpt-oss or Llama, you can run it on servers you control and pay for the servers instead, busy or not. For most projects the API costs less, both to build and to run. The section on running your own model below shows when your own servers pay off.
A custom model for predictions
When the task is to predict something from your data, such as demand, churn, or the risk that a payment is fraud, you need a machine learning model trained for that one prediction, not a language model. At Ronas IT, such a model starts at $15,000 and takes at least 6 weeks.
Training a new language model of your own is almost never worth it for a business that uses AI rather than sells models. A 2024 study by Epoch AI found that the cost of training frontier AI models had grown 2 to 3 times a year for eight years, which suggests that the largest models will cost over a billion dollars by 2027. A company that uses AI pays for those models through usage fees instead.
What AI costs after launch
After launch you pay up to three things every month: the model provider's usage fee, hosting for your system, and support if you choose a plan. At Ronas IT, support for an AI app starts at $1,000 per month. The usage fee is the one that grows with traffic, so estimate it before the build starts.
Providers such as Anthropic, OpenAI, and Google price usage per million tokens, which are small pieces of text. Text the model writes costs more than text it reads, and each provider's most capable models cost many times more than its smallest ones.
Take an assistant that answers 1,000 requests a day. Each request sends about 2,000 tokens of instructions and context, and each answer is about 300 tokens long. Over 30 days, that is 60 million input tokens and 9 million output tokens. At list prices in November 2026:
| Model | Price per million tokens, input / output | Usage per month |
|---|---|---|
| Claude Haiku 5.5 or GPT-6 Luna | $0.10 / $0.50 | About $11 |
| Gemini 3.8 Flash, from January 1, 2027 | $1.50 / $7.50 | About $160 |
| Claude Sonnet 5.5 or GPT-6.2 Sol | $2 / $10 | About $210 |
| Gemini 4 Carbon, from December 1, 2026 | $4 / $20 | About $420 |
| Claude Fable 5.2 or GPT-6.2 Astra | $8 / $40 | About $840 |
The same requests cost about 80 times more on the most capable models than on the smallest ones. The table is a first estimate, not a quote: each provider splits text into tokens differently, so the same request can count as a different number of tokens. Google has also announced that both Gemini prices in the table double from the dates shown. Prices change often, and in both directions: Anthropic released Claude Fable 5.2 in November 2026 at $8 and $40, below the $10 and $50 of Claude Fable 5.1.
Running your own model
An open model on your own servers replaces the usage fee with a server bill that runs around the clock. According to its model card, OpenAI's open model gpt-oss-120b fits on one GPU with 80 GB of memory, such as an NVIDIA H100. A larger model, such as Llama 4 Maverick, needs a whole server with eight H100 GPUs. On AWS, such a server, p5.48xlarge, costs about $55 an hour at on-demand prices in November 2026, or about $40,000 a month if it runs all the time.
That bill is only the start:
- Failover: one server is one point of failure, so a live system needs a second one or a fallback to an API.
- Idle time: you pay for the server at night and on weekends, when the load is low. An API charges only for the requests you send.
- People: engineers have to install new versions of the model, watch the load, and add servers when traffic grows.
- Quality: the open models you can run on one server can do worse on hard tasks than the most capable models behind an API, so test them on your own examples first.
Compare that with the example above: the same assistant costs about $210 a month on a mid-range model through an API. Your own servers pay off at a volume many times higher and steady through the day, or when your data may not leave servers you control. Reserved or long-term server contracts lower the hourly price, but they also lock you into a model size before you know the load.
Costs that follow the provider's schedule
Two later costs rarely appear in a quote, because they follow the provider's schedule rather than your traffic. The first is the retirement of a model. Anthropic publishes a retirement date for each deprecated model and gives at least 60 days' notice for publicly released models; after that date, requests to the old model fail. Moving to the replacement is project work: the AI has to pass your test examples again on the new model, and instructions written around the old model's behavior may need changes. The usage fee can change too, even when the price per token stays the same. When Anthropic released Claude Opus 4.7 in April 2026 at the price of Opus 4.6, it noted that the new model's updated tokenizer can turn the same input into roughly 1.0 to 1.35 times as many tokens, depending on the content. After each model change, run the usage estimate again on your own requests.
The second applies to RAG. The search index stores each passage of your documents as an embedding, a list of numbers produced by one particular embedding model, and a question can only be matched against passages embedded by the same model. Google's documentation for its Gemini embedding models calls the embedding spaces of two of its versions “incompatible” and says an upgrade means re-embedding all existing data. Switching the embedding model, by choice or because the old one is retired, therefore means processing your whole document collection again and paying for that run. Ask the vendor to list every model the system depends on and to explain how a change of each one would be tested and priced.
How to keep AI development cost down
An AI project has two bills: the build, which you pay once, and the usage fee, which you pay every month for as long as the system runs. Many of the savings on the usage fee are decided in the build.
Keep the monthly usage bill down
- Choose the model the task needs, not the most capable one: run your test examples on a smaller model first and move up only if it fails them. Split the work: a small model can sort requests and pull fields out of a form, and a more capable one handles only the hard steps. In the table above, that is the difference between about $11 and about $840 a month for the same number of requests.
- Send the model only the context it needs: you pay for every token in every request. Keep the instructions short, let RAG return only the few passages that answer the question, replace a long conversation history with a short summary, and give each step only the tools it uses. Anthropic's engineers describe the goal as “the smallest possible set of high-signal tokens.”
- Do in code what code can do: calculations, date checks, format validation, lookups in a database, and rules with a fixed answer should run as ordinary code. Code costs nothing per request, runs faster, and gives the same result every time. The model handles only the part that needs language or judgment. When an agent repeats the same sequence of steps for every case, turn that sequence into a script.
- Do not pay twice for the same text: providers charge a fraction of the input price for instructions read back from their prompt cache, so keep the fixed part of each request first and unchanged. Store answers to questions that come up again and again. Send requests that can wait, such as a nightly report, through the batch API, which Anthropic, OpenAI, and Google offer at half the price.
- Set limits for agents: an agent sends its growing context again at each step, so a task that takes twenty steps costs far more than one request. Set a maximum number of steps and tokens per task, and stop the agent when it reaches them.
- Measure the cost of each feature: without numbers per feature, you cannot tell which change saved money.
“Log the model, the tokens, and the cost of every request from the first day, split by feature. When the bill grows, you want to know which feature grew and why, not just the total from the provider.”
Alexander Storozhevsky, Lead developer at Ronas IT
Keep the build cost down
Spend first on the work that tells you whether the rest is worth building: a written scope and, when the task is uncertain, a test on your own examples. Then build one task from start to finish before you add the next one.

- Check the tools you already pay for: if the AI feature of your CRM or help desk does the task, you pay for its plan or usage and no development. If a ready tool fits, we will tell you.
- Start with one task: pick one task, the people who do it now, and the number that you want to change, such as the time spent on each ticket. Our guide to adding AI to existing business processes shows how to find that step.
- Write the scope down before the build: if you have none yet, the analysis phase, at $1,000 to $2,000, ends with a written scope, a budget, and a timeline, so the build is priced on facts.
- Test before you commit: a proof of concept run on your own examples during the analysis phase shows whether a model handles the task before you pay for the build.
- Split the work with your own team: on an AI tool for a creative agency, one business analyst and one designer from our team did the analysis and designed every screen of the app. Together, the analysis and the design took four months and cost $16,000+. The analysis narrowed a long feature list to the features the agency needed most. The client's own developers then took on the coding and the AI integration.
How to choose AI development services by comparing quotes
Compare what each quote includes, not only the total. Two totals that differ by half may describe different projects.

- Approach: which model or approach the company proposes, and why a simpler one would not be enough.
- Acceptance: the set of anonymized or synthetic examples with expected answers that the result must pass, agreed before the build.
- Running costs: what you will pay after launch, such as the model provider's usage fee, hosting, and support, listed apart from the build.
- Separate lines: what is inside the price and what is billed on top, such as project management, design, and new APIs.
- Support after launch: what the plan covers and how fast the team responds, both in writing. At Ronas IT, support terms and response times are agreed in the support agreement.
- Contract model: a fixed price for a written scope, or time and materials while the approach is still open. Our article on fixed price vs time and materials compares the two.
- Ownership: whose cloud account the system will run in and how you receive the code. Our projects start in your own cloud account.
What to do next
- Write down the one task for AI, who does it now, and the number that should change.
- Check whether software you already pay for has an AI feature that does this task.
- Estimate the monthly usage fee as in the example above: requests per day, times the tokens per request, times the provider's price, counted for input and output separately, then times 30 days.
- Find your line in the price table. If you have no written scope yet, start with the analysis phase.
Frequently Asked Questions (FAQs)
How much does it cost to build an AI system?
At Ronas IT, an AI chatbot, agent, or other AI feature inside your product or process starts at $3,000 per feature, a new product built around AI starts at $5,000 as a basic MVP, and a custom machine learning model starts at $15,000. Each price already includes design, project management, and testing.
How much does AI cost per month after launch?
You pay the model provider for usage, pay for hosting, and pay for support if you choose a plan. Usage depends on traffic and the model. At list prices in November 2026, an assistant that answers 1,000 requests a day, with about 2,000 input tokens and 300 output tokens per request, costs about $11 a month in usage on Claude Haiku 5.5, about $210 on Claude Sonnet 5.5 or GPT-6.2 Sol, and about $840 on Claude Fable 5.2 or GPT-6.2 Astra. Our support for an app starts at $1,000 per month, and for a whole system of several apps at $2,000.
How much do AI tokens cost?
Providers set a price per million tokens, and text the model writes costs more than text it reads. In November 2026, a million tokens of Claude Sonnet 5.5 or GPT-6.2 Sol cost $2 to read and $10 to write, Claude Fable 5.2 and GPT-6.2 Astra cost $8 and $40, and the smallest models, Claude Haiku 5.5 and GPT-6 Luna, cost $0.10 and $0.50. Google raises the price of some Gemini models on set dates, so check the provider's current prices before you plan a budget.
Is it cheaper to use an AI API or to build your own model?
For text tasks, a ready model through an API is cheaper to build and to run: you pay for the integration and then per use, with no training and no servers. An open model on your own servers needs GPU servers that you pay for by the hour whether they are busy or not, and a server for a large model can cost about $40,000 a month. A custom model makes sense if you need to predict something from your own data, for example demand or churn. At Ronas IT, such a model starts at $15,000.
How much does it cost to run your own AI model?
You pay for the servers instead of a usage fee. A smaller open model, such as gpt-oss-120b, fits on one GPU with 80 GB of memory. A larger one, such as Llama 4 Maverick, needs a server with eight H100 GPUs, which costs about $55 an hour on AWS at on-demand prices in November 2026. Add a second server for failover, and the engineers who update, monitor, and scale the model. This pays off only at a high and steady volume, or when your data may not leave servers you control.
How much does an AI chatbot cost?
At Ronas IT, an AI chatbot or assistant for your website, app, or messenger is one AI feature: it starts at $3,000 and takes at least 1 week. Its connections to your systems that already have an API, such as a help desk where it creates tickets, are part of that work, and so are design and project management. An agent that chooses its own steps across your systems is priced the same way, per AI feature. A system with no API needs one first, from $2,500.
How long does AI development take?
A chatbot or another AI feature takes at least 1 week, a new AI product at least 2 weeks as a basic MVP or at least 5 weeks as an extended MVP, and a custom model at least 6 weeks. An optional analysis phase before the build takes at least 1 week. On our projects, each sprint usually lasts one to three weeks and closes with a result you can check.
What are the hidden costs of AI development?
A quote may not show hosting, support after launch, work to collect data scattered across spreadsheets and mailboxes, an API for a system that has none, or the model usage fee, which grows with traffic. Ask for each of them as a separate line in the quote. Two costs are not on this list: the design of new screens, included in the price of the AI feature, and preparing the anonymized or synthetic data a custom model is trained on, included in the price of the model. One cost comes later: when a provider retires the model you use, moving to its successor needs a new round of tests, and a retired embedding model means embedding every document passage again.
Should an AI project be fixed price or time and materials?
A fixed price works when the scope is written down and the approach is settled, for example after an analysis phase or a proof of concept. Time and materials suits work where the approach can still change after testing on your data, such as choosing between a prompt, document search, and fine-tuning.






