How to get a working AI prototype in weeks, not quarters

An AI product costs three things: the build, the model usage that grows with every request, and the upkeep after launch. The cheapest way to learn whether the full spend is worth it is a prototype that tests one assumption in weeks. Reveneau builds this way: AI writes all of its code, every change has to pass an eval suite written from the specification, and because it uses AI instead of hiring more engineers, that saving goes into the client's price.
If an AI idea has a flaw that ends it, the flaw is there on the first day. A team that waits six months for its first demo finds the flaw after six months of budget is spent. A smaller first build finds it in a week or two. So show an early, unfinished version to real people, see what breaks, and change direction while that is still cheap.
How much does it cost to build an AI product?
The cost has three parts: building it, paying the model provider for every request it handles, and keeping it working after launch. Only the second part has a public price list, so a fair estimate starts there and treats the other two as work to scope.
Building it. This is the work of turning a working idea into something people can trust: the interface, the connections to your other systems, the handling of wrong or missing answers, and the tests that check each change. The price depends on how many systems it touches and how often it is allowed to be wrong. Our guide on what it costs to build an AI product goes through each of these.
Paying for the model. Model providers charge per token, the small pieces of text a model reads and writes. On Anthropic's pricing page, read on 5 November 2026, Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens. Claude Haiku 5.5 costs $0.10 and $0.50 for prompts up to 100,000 tokens. Requests sent as a batch cost $1 and $5 on Sonnet 5.5. Here is what that means for one request that sends 2,000 tokens and gets 500 back (sizes chosen for this example):
| Model | One request | 100,000 requests a month |
|---|---|---|
| Claude Sonnet 5.5 | $0.009 | $900 |
| Claude Haiku 5.5 | $0.00045 | $45 |
For the same traffic, the bill is $900 on one model and $45 on the other. Whether the smaller model is good enough for your task is one of the first things a prototype should test, along with how many tokens a real request uses.
Keeping it working. Providers retire models and change prices. The same Anthropic page lists four older Claude models as retired, each still offered on Google Cloud, and two of them also on Amazon Bedrock. When the model under your product changes, its checks have to run again on the new one, and someone has to fix what fails. Budget for that as a cost that repeats. Our guide on how to evaluate an LLM's answers covers the checks themselves.
Start with the assumption that can end the idea
Every AI product idea depends on one assumption that matters more than the rest: that users want this automated, that the model can handle the messy real version of the task, that the cost per request works at the volume you need. Find that assumption first, and build only enough to test it. Everything else can wait.
Do not treat every part of the idea as equally uncertain. In a support automation tool, the chat interface is well understood work and carries little risk. The risk is whether the model can read a real customer message, full of typos and half-finished sentences, and route it correctly often enough to be trusted. That is the assumption worth testing on day one. So build a reduced version that does only that: it takes a real ticket, produces a routing decision, and shows its reasoning. No login, no dashboard, no settings page. If the routing is wrong half the time, you have learned the most important thing about the idea in a week, for the cost of a week.
Write the critical assumption down as a plain sentence before you build anything. If you cannot name it, the idea is not scoped yet, and more planning meetings will not change that. A prototype can only test a question you have written down.
Use the model as it is
Start with what a ready-made model from the newest generation can already do. Invest in fine-tuning only once you know the idea is worth the extra cost and work.
Starting with fine-tuning is tempting, because it feels like the serious, technical thing to do. In practice it slows you down at the moment speed matters most. Fine-tuning needs clean training data, an evaluation setup, and time you do not have yet, and it commits you to decisions before you know whether the idea works at all. We would rather spend that week testing the concept than improving a model for a concept that might fail as soon as users try it.
The newest models handle a wide range of tasks well enough to prove or disprove an idea without custom training. When a model is not good enough, try better prompting first, then give it the right context, then split one hard step into two smaller ones it can handle. Those changes take hours. Fine-tuning takes weeks. Save it for the point where the idea is proven and you are improving a known-good feature for cost or accuracy.
Show it to five real users before fifty
A polished demo to fifty polite people teaches you little. Five real users who try to use the thing, and try to break it, show you the problems that matter sooner.
A demo controls the inputs, and real users do not. When you run the demo, you give the model the clean examples it handles well, and everyone agrees. Five real users will paste in the unclear, half-formed inputs that fill their working day, and those are the inputs the demo never tested.
Keep this round small on purpose. Five people is enough to see the same failure appear more than once, which shows a pattern, and it is small enough that you can sit with each person and watch. Do not send a link and wait for feedback. Watch where they hesitate, where they retype something, where they stop trusting the output. Those moments tell you more than any survey. Fix what you see, then show it to more people. Fifty users is the round for confirming a result you already have.
Skip production infrastructure until the idea proves itself
Monitoring, alerts, and cost controls are essential once an AI feature runs for real users at scale. A prototype answers one question, whether the idea is worth building for real, and it can answer that question without them. The test that answers it is itself a small evaluation: a set of real inputs, each with the answer it should get. Keep that from the first day. The production layer around it waits.
This is hard for good engineers to accept, because building the infrastructure feels like doing the job properly. But a monitoring dashboard for a feature nobody has agreed to keep is work spent on a decision you have not made yet. The prototype has one job, which is to answer a yes or no question about the idea. Everything that does not move you toward that answer is a delay that only looks like careful work.
Without that layer the prototype is unsafe to run at scale, and you have to say so to everyone who uses it. A prototype is a learning tool, and it should never quietly turn into a product. The moment the idea proves itself, monitoring and cost controls become the first thing you build. Building them in that order means you never make something ready for production that you are about to throw away.
Why a small first build saves time and money
A small first build is the fastest way to find out whether an idea deserves the full investment, before that investment is made.
A narrow scope also protects the good ideas. When the first build is small, stopping a weak idea costs a week, so nobody feels the need to defend a bad plan because a lot of money went into it. And when an idea does prove itself, you reach the full build already knowing the hard part works, so the larger spend goes toward something you have tested.
How Reveneau runs this
Our AI development work starts with this kind of scoped prototype. We name the critical assumption first, build only enough to test it, show it to a small number of real users, and let the result decide what happens next. AI writes the prototype's code, and every change to it passes the eval suite written from the specification, as on every Reveneau build. Monitoring, alerts, and cost limits are added once real use justifies them.
Because AI writes the whole implementation, the build is no longer the slowest step. That is why the first days of a prototype go on scope rather than architecture. When building is quick, the only way to be slow is to build the wrong thing.
Related guide: How to build an AI product.
Sources
- Anthropic, "Pricing", Claude Platform documentation (read 5 November 2026): https://platform.claude.com/docs/en/about-claude/pricing


