How Much Does It Cost to Maintain an AI-Built SaaS After Launch?

The monthly bill for a small AI-built SaaS is rarely the scary part. In the three worked budgets below, the tools, hosting, email, payments and AI model usage come to about $50 a month with 100 users, about $100 with 1,000 users, and about $2-300 with 10,000 users. The expensive part is the work behind the bills: fixing bugs, updating dependencies, reviewing security and keeping an AI feature from breaking when a model changes. Once you price that in, the same three budgets reach roughly $150, $1,460 and $7,660 a month.

This guide shows where each number comes from, using current public prices for Lovable, Bolt, Vercel, Supabase, Resend and the major AI model providers. The examples are illustrative, because your usage will differ, so the guide also gives you formulas to replace my assumptions with yours. It covers the tool layer, hosting, AI feature costs per user, payments, the people cost most founders leave out, why AI-written code tends to need more upkeep, the bills that surprise people, and when it is cheaper to rebuild than to patch.

The short version: three budgets

All three examples assume a B2B tool with a $29 monthly plan and an AI feature that users call 20 times a month. They use the prices in the sections that follow. Lines marked as assumed are placeholders you should replace with a quote or your own usage data.

Stage 1 assumes you do all the maintenance yourself, so your time is not priced. Stage 2 assumes about 8.7 hours a month of paid help at $75 an hour plus $2,400 a year for an outside security review. Stage 3 assumes 40 hours a month at $75 and a $5,000 annual review. The $75 rate and the review budgets are placeholders, not market quotes.

Two patterns show up. Running costs stay well below a quarter of revenue at every stage, which is healthy for software. The share only climbs toward a third when you add people, and that is where founders get surprised. The next sections take the layers one at a time.

What the budgets mean per customer

Two calculations turn those totals into decisions.

All-in cost per paying customer. Divide the all-in monthly cost by paying customers: $152 ÷ 20 = $7.60 at Stage 1, $1,464 ÷ 150 = $9.76 at Stage 2, and $7,659 ÷ 1,200 = $6.38 at Stage 3. Against a $29 plan, each customer leaves roughly $19 to $23 after the cost of serving them and maintaining the product, before marketing, tax and your own pay.

Customers needed to cover the fixed costs. Take the costs that do not rise with each new customer, which are tools, hosting, maintenance and security review, and divide by what each paying customer contributes after payment fees and their share of AI usage. At Stage 2 the fixed costs are $221 + $652 + $200 = $1,073. A paying customer contributes $29 − $1.34 in fees − $1.26 in AI usage, which is $26.40. So $1,073 ÷ $26.40 means about 41 paying customers cover the fixed costs, and the other 109 contribute to profit and growth. At Stage 1 the same method gives about 4 customers.

The free-user trap hides inside these numbers. At Stage 2, 850 of the 1,000 users do not pay, yet if they make the same 20 AI requests each, they generate $161 of the $189 AI bill, or about 85% of it. A free plan with an AI feature is a marketing cost you set the size of, so give free users a smaller allowance than paying users.

Layer 1: the tool layer, and whether to keep paying it

The first line on most bills is the AI builder or coding assistant you used to create the product. As of the pricing guides checked in November 2026, Lovable's Pro plan is $25 a month with 100 monthly credits plus 5 daily credits, and its Business plan is $50 with SSO and team controls. Bolt's Pro plan is $25 a month for 10 million tokens, with Teams at $30 per member. Cursor's Pro plan is listed at $20 a month.

After launch, these subscriptions change meaning. You are no longer buying a way to build an app, you are buying a way to fix and extend one. Two things follow:

• Fixing bugs spends credits, too. Both Lovable and Bolt meter usage, and debugging sessions with a large codebase burn more than a fresh prompt. Budget for a higher tier in months when you ship a lot, and watch how fast credits drain in a repair session.

• Pay for one tool, not three. Many founders keep a builder, a coding assistant and a chat subscription running at once. Pick the one you actually use to change the product, and cancel the rest. The guide to cutting startup subscription costs walks through the audit.

If you export the code to GitHub and work in an editor from then on, you may not need the builder subscription at all. Do that only if someone on your team can read and change the code, which is a bigger question than the $25.

Layer 2: hosting and database

Hosting is where the pricing page and the real bill diverge most, so start with the rules that can bite you.

Vercel. The Hobby plan is free but limited to personal, non-commercial projects, so a product that charges customers belongs on Pro. Pro costs $20 per developer seat per month and includes $20 of usage credit per seat, plus 1 TB of data transfer and 10 million edge requests, according to current pricing summaries. Viewer seats are free. Beyond the included amounts, bandwidth, function execution, image optimization and builds are billed by use, and published overage rates differ by source and region, so read the live pricing page. One calculator example puts a three-seat account with 250,000 monthly visitors and 5 million API requests at about $305 a month, which is the kind of jump to expect when a product gets traction.

Supabase. The Free plan includes two active projects, 500 MB of database storage and 50,000 monthly active users, and pauses projects after a week of inactivity. A paused database means a broken product, so a live SaaS belongs on Pro at $25 a month per organization, which includes a $10 compute credit. Larger compute adds to that, which is why Stage 2 assumes $25 extra and Stage 3 assumes $175 extra. Those two figures are my assumptions.

Neon and Upstash. Neon's paid plans are usage-based with no monthly minimum, and Upstash Redis has a free tier of 256 MB and 500,000 commands a month, then pay-as-you-go at $0.20 per 100,000 commands. The comparison of Neon, Supabase and Upstash on Vercel explains how to choose between them.

Region and connection mistakes also cost money. A database far from your functions adds latency to every query, and serverless functions that open a new connection on each request can exhaust limits and trigger upgrades you did not need. Both are cheap to fix and expensive to ignore.

Layer 3: the AI feature, priced per user

If your product calls a language model, this line grows with usage, not with the number of customers, and it is the one most likely to damage your margin. The method is the same every time:

Cost per request = (input tokens × input price + output tokens × output price) ÷ 1,000,000

Published price tiers from the major providers sit roughly in three bands: small models near $1 per million input tokens and $5 per million output tokens, mid-size models near $3 and $15, and flagship models near $5 and $25. Top-tier models cost more, and OpenAI's lineup spans a similar range. Prices move with each release, so check the provider's pricing page before you rely on these.

The table applies those bands to three kinds of request. A light request is a short chat or summary with 2,000 tokens in and 500 out. A retrieval request pulls documents into the prompt, at 8,000 in and 800 out. An agent task chains many steps, at 30,000 in and 3,000 out.

Multiply by usage and the picture changes. A paying user who makes 20 light requests a month costs $0.27 on a mid-size model, which is nothing. The same user making 100 agent tasks a month on a flagship model costs $22.50, which is 78% of a $29 plan before payment fees. That is the trap: an AI feature that looks cheap in testing can erase your margin once a few heavy users arrive.

Three levers work reliably:

• Prompt caching. Providers discount repeated input. Anthropic's published rate for cached input is up to 90% lower. If 75% of a retrieval request's input is cached, the mid-size cost drops from $0.036 to $0.0198, a saving of 45%. The saving depends on whether your prompts really repeat and whether cached content stays warm.

• A smaller model for easy steps. Classification, extraction and routing rarely need a flagship model. Moving them to a small model cuts that step's cost by about two thirds.

• Batch processing for work that can wait. Anthropic's batch API halves input and output prices for asynchronous jobs.

Then add controls: a hard cap on tokens per request, a monthly usage allowance per plan, and an alert when daily spend jumps. A runaway loop in an AI feature is one of the few ways a small SaaS gets a four-figure bill overnight.

Model upkeep is also a recurring job. Providers retire older model versions on published schedules, so plan a few hours each time you migrate, and keep a small set of test prompts that tells you whether answers got worse.

Layer 4: payments

Payment fees are a cost of selling, not of maintaining, but they belong in the budget. Stripe charges 2.9% plus 30 cents per card payment, with 1.5% more for international cards, 1% for currency conversion and roughly 0.5% to 0.7% for Stripe Billing. On a $29 plan with the base rate and Billing, that is about $1.34 per payment, or 4.6%. Merchant-of-record platforms like Paddle charge about 5% plus 50 cents and handle sales tax and invoicing, which matters if you sell across borders.

At 150 paying customers on a $29 plan, Stage 2 pays about $202 a month in fees. On a $5 plan the same fixed 30 cents takes about 9.6%, so cheap plans lose more. The guide to payment gateways compares the options in more detail.

Layer 5: the small services that add up

Each of these is cheap alone and meaningful together:

• Email. Resend's free tier covers 3,000 emails a month with a 100-a-day cap. Pro is $20 a month for 50,000 emails and $90 for 100,000. A product with a few hundred new users a month hits the free cap quickly.

• Error monitoring. Sentry's Team plan is listed at about $26 to $29 a month depending on the guide. Without monitoring, you learn about outages from customers.

• Analytics. PostHog's free tier covers 1 million events a month, which is enough for most early products.

• Authentication. Recent pricing guides show Clerk's free plan covering 50,000 monthly retained users and Pro at $25 a month. Older guides list lower limits, so check the current page. If you use Supabase Auth, this is included.

• Domain and workspace. A domain is about $1 a month, and a team email suite is about $6 per user.

One example stack from a published startup guide totals about $52 a month for a domain, Vercel Pro, Supabase Pro and a workspace seat, before AI usage, payments and any people.

The line most budgets leave out: maintenance work

Software needs regular care even when nothing is broken. For a small SaaS, the recurring jobs look like this:

The hours column is a planning assumption, not a measurement, so treat the total of about 8 to 15 hours a month as a starting point and track your own. At $75 an hour that is $600 to $1,125. If you do the work yourself, the cost is your time. If you are a solo founder with a full-time job, ask what else those hours could earn or what that time is worth to you.

Why AI-built code tends to cost more to maintain

This is the part that changes the budget compared with a conventionally built product. The evidence is mostly from vendors and researchers measuring AI-generated code in general, not Lovable or Bolt apps specifically, and several of the sources sell security services, so treat the numbers as signals.

• Veracode's analysis of over 100 models found that roughly 45% of AI-generated code samples introduced security flaws from the OWASP Top 10, with the rate not improving across testing rounds.

• CodeRabbit's December 2025 analysis reported that AI co-authored code shipped about 1.7 times more major issues and 2.74 times more security vulnerabilities than human-only code.

• GitClear's analysis found the share of code that is refactored fell from about 25% in 2021 to below 10% in 2024, which means more copy-paste and less cleanup.

• Georgia Tech's Vibe Security Radar tracked CVEs attributed to AI-generated code rising from 6 in January 2026 to 15 in February and 35 in March.

• A scan by Escape.tech and Wiz of 5,600 vibe-coded applications found over 400 exposed secrets and 175 instances of exposed personal data.

None of this proves your app is insecure. It tells you where to look. In practice, the maintenance burden shows up in four places:

1. Access control. Missing or weak row-level security and authorization checks are the classic flaw. Test that one user cannot read another's data.

2. Secrets. Keys committed to code, pasted into prompts, or exposed in the browser. Rotate any key that has ever been in a prompt or a repository.

3. Webhooks and payments. Unverified payment webhooks and trust in client-side prices.

4. Code you do not understand. If nobody can explain a module, every change is a risk, and every bug takes longer to find.

Tools help but do not replace review. Lovable, for example, offers a built-in security scan that can flag and fix some issues, and static analysis, secret scanning and dependency checks can run in your build pipeline. Add tests for the paths that matter most: sign-up, payment, permissions and data export. If you plan to sell the product later, this clean-up is also what buyers look at, as the guide on selling a SaaS built with Lovable or Bolt.new explains.

Bills that surprise people

• Running a paying product on a free plan. Vercel Hobby forbids commercial use, and Supabase Free pauses a quiet project after a week. Both break a live product or the terms.

• Usage overages. A traffic spike, a chatty API or a crawler can add bandwidth and function charges. Set spend alerts on every platform that offers them.

• A runaway AI loop. A bug that retries a model call endlessly can cost hundreds of dollars before you notice. Cap tokens, calls per user and daily spend.

• Credit burn in the builder. A long repair session can use a month of credits in a day.

• Annual renewals. Domains, plans and add-ons renew at a higher price than you remember.

• Email limits. Hitting a sending cap stops password resets and receipts, and customers experience that as an outage.

• Compliance and policy work. Privacy policy updates, data processing terms and tax registration cost time or professional fees. If you serve EU users with an AI feature, check whether new transparency obligations that began applying in August 2026 affect you, and confirm with a lawyer.

• Refunds and chargebacks. Payment disputes carry fees, so keep a small reserve.

Rescue or rebuild: the expensive decision

At some point a patched prototype can cost more than it saves. Community reports cited by agencies peg a professional rebuild of a vibe-coded application at $5,000 to $30,000, and one agency that does this work says rebuilds run three to five times the cost of a targeted rescue. Those are vendor figures and rough ones. They are useful for scale, not for quotes.

Rescue is usually the better first move when the data model is sound, the main flows work, and the problems are concentrated in security, tests and documentation. Rebuild becomes the better option when changes keep breaking unrelated features, nobody can explain core modules, the data structure blocks features customers are asking for, or each fix takes longer than the last. A practical test: ask a developer who has not seen the code to estimate a small change. If the answer is a long investigation before any work, that is the signal.

How to keep the bill down

• Use free tiers only where terms allow. Free plans suit staging and prototypes, not paying customers.

• Pay for one builder or assistant. Cancel overlapping tools.

• Cache and downsize AI calls. Caching can save about 45% on retrieval-style requests in the example above, and small models handle easy steps.

• Put usage limits in your plans. Include a monthly allowance and sell more as credits. This protects margin from heavy users.

• Choose a payment setup that matches your plan price. Fixed per-transaction fees hurt low-priced plans.

• Review monthly. Twenty minutes with your invoices catches most surprises.

• Automate the boring checks. Dependency updates, secret scanning and tests in your build pipeline cost little and catch problems early.

• Share the load. The AI tools guide for indie hackers lists tools that remove solo work, which can cut maintenance hours.

Build your own budget

Use this formula, and fill each line from a real invoice or usage figure:

All-in monthly cost = tools and hosting + AI usage + payment fees + maintenance hours × hourly rate + security review ÷ 12

Then track three numbers every month. First, running cost per active user, so you see whether costs scale with usage. Second, gross margin after hosting, AI usage and payment fees. In Stage 2 above, those lines total $531 against $4,350 of MRR, a margin of about 88%. Third, maintenance hours, because a rising number is an early warning that the code is becoming harder to change.

A simple decision rule follows. If gross margin drops below about 70%, look at AI usage and plan limits first. If maintenance hours keep climbing, schedule a code review before adding features.

FAQ

What is the minimum monthly cost to keep an AI-built SaaS online? For a product with paying customers, the floor is roughly Vercel Pro ($20), Supabase Pro ($25) and a domain, about $46 a month, plus payment fees and any AI usage. Free tiers can cover a prototype, but they come with pauses, limits and in Vercel's case a ban on commercial use.

Do I need to keep paying for Lovable or Bolt after launch? Only while you use it to change the product. Credits and tokens are consumed by every edit and repair. If you move to editing the code directly, you can drop the subscription, as long as someone can maintain the code.

How much does an AI feature cost per user? It depends on the request size, model tier and usage. A light request on a mid-size model costs about a cent and a half, so 20 a month costs about 27 cents, while 100 agent tasks a month on a flagship model costs about $22.50. Calculate it from your own token counts.

Is an AI-built app more expensive to maintain than a traditionally built one? The research suggests more security issues and less refactoring in AI-generated code, which tends to raise review and cleanup work. Most of the evidence is not specific to Lovable or Bolt, so test your own app before you assume the worst.

How much should I budget for a security review? Get quotes, since prices vary widely by scope. The examples here set aside $2,400 to $5,000 a year as a placeholder, which would buy roughly 32 to 67 hours at $75 an hour.

When should I rebuild instead of patch? When small changes keep breaking other features, nobody can explain core code, or the data model blocks what customers need. Compare a targeted rescue estimate with a rebuild estimate before deciding.

Start with one hour and your last three invoices. List every recurring charge, add your payment fees and any AI usage, and write down how many hours you spent on fixes last month. That gives you a real all-in figure, and the formula above shows where to cut first.