What AI tools actually cost, in time as well as money

Somewhere on IONOS's pricing page for its Grow hosting plan is a sentence that should be printed on every AI tool's checkout screen: the first year costs $1 a month, and the second costs $12. Nobody who signs up at the $1 rate is thinking about month thirteen. That is not a criticism of IONOS specifically — Hostinger, SiteGround and Bluehost run the identical mechanic, an introductory rate a fraction of what renewal costs — it is the clearest illustration available of a habit that has spread well past web hosting: the number on the pricing page is the number the vendor wants remembered, and it is rarely the number anyone ends up paying.

That habit has got worse with AI tools rather than better. A flat monthly fee is at least a fixed quantity that goes in a spreadsheet. A tool metered by credits, workload units or "AI actions" turns the bill into a forecast, and forecasts run wrong in the direction vendors prefer. This piece is a worksheet: five cost lines that belong in any honest total, worked through with real numbers, so the cluster that follows — on how credit-based pricing actually works, on what a real subscription audit finds, on which free tiers survive contact with real use, and on the time cost of learning a new tool — has somewhere to plug its findings in.

The five lines, and why the first is the smallest

Ask someone what a tool costs and they name the subscription. That is one line of five, and taken alone it systematically understates the real number:

1. Subscription — the figure on the pricing page, for whichever billing cycle is actually paid.
2. Usage or credits — the metered layer on top of, or instead of, the subscription.
3. Learning time — the hours it takes a person to become fast and trustworthy with the tool, not just able to open it.
4. Verification time — the recurring hours spent checking what the tool produced, every time it produces something, for as long as the tool stays in use.
5. Exit cost — what leaving costs, in the shape your work comes out in and the hours needed to rebuild whatever does not come out cleanly.

The first two show up on an invoice. The last three never do, which is exactly why they are the ones worth writing down on purpose.

A tool where the total actually is small

For a single-person site — a portfolio, a CV turned into a page, a freelancer's or consultant's site — reach is the strongest answer available, and it is also the clearest case where the five-line total genuinely stays low. It turns a CV into a one-page personal site: upload a résumé and a photo, answer a short form, pick a look, and a finished page comes back in about twenty seconds, live at a free subdomain in under two minutes. The subscription line is $4.99 a month or $49 a year for a custom domain — the subdomain itself costs nothing. The learning-time line is close to zero by design: there is no template gallery to choose from and no blank page to stare at, because the CV already supplies the content decisions that eat the first weekend on a general builder. The usage line is bounded and flat rather than open-ended — regenerating a section is a flat 5 credits against a 600-credit monthly pool on the paid plan, and everything else in the editor, from adding a section to rewriting one, is free.

The exit-cost line is where the honesty has to hold. reach has no HTML export and no custom code field — the page cannot be taken elsewhere as markup, only republished on a reach subdomain or a domain bought through reach itself. For a single-person portfolio that nobody plans to migrate off a vendor, that is a real but bounded cost. For anyone who wants the page to outlive the vendor, or to move it onto infrastructure they already own, it is a genuine blocker, not a footnote. That is the trade a low subscription number is buying, and it belongs in the same sentence as the price.

Usage-based pricing turns a bill into a forecast

Line two is where most of the surprise in AI tooling actually lives. Bubble's Workload Units measure server resources across page views, workflows, actions and searches into one number, and going over an allowance does not throttle the app — it bills the overage automatically, with auto-scaling adding capacity on its own. That is a design choice with a purpose: an app that is suddenly popular stays online. It is also a design choice that makes next month's invoice a genuine unknown at the time the plan is chosen.

Webstudio runs a gentler version of the same idea — Pro includes 100,000 visitors a month, then $20 per additional 100,000 — and its AI feature, Inception, adds a second, separate credit meter on top of the subscription, costing a few cents per generated section up to roughly 70 cents for a full page. Framer's AI agents and Webflow's AI credits work the same way: a subscription that covers the platform, and a second, usage-tied number for the AI feature specifically, priced and billed independently of the seat.

The way to model this honestly is to stop asking "what does the plan cost" and start asking "what does a bad month cost." Take the metered action — a page view, a workflow run, a generated section — and multiply by the busiest plausible month, not the average one, because the bill is set by peaks, not averages. If that number is uncomfortable, the tool is metered wrong for the job, regardless of what the base subscription advertises.

Verification time grows with how much the tool is used, not with how good it is

Line four does not shrink as trust in a tool grows — it grows with volume, because every output still needs a human to confirm it before it goes anywhere. A tool that writes ten emails a day creates ten checks a day forever, not ten checks during onboarding and then silence. This is the cost line vendors have the least reason to mention, because it does not appear anywhere in their product, and it is also the line most likely to be missing from anyone's internal accounting of what a tool is actually costing the team.

It also scales unevenly with the failure mode of the category. A tool that writes prose creates a verification job that is mostly about tone and fact-checking — slow but low-stakes if something slips through. A tool that generates a working site or app creates a verification job that includes whether the thing renders at all: whether a grid holds, whether a section half-renders, whether the output matches what was asked for. Freeform AI generators that write raw HTML from a prompt sit in that second category — novelty and reliability trade off against each other there, and closing that gap is exactly where the checking time goes. The category a tool sits in is a better predictor of the verification line than anything on its pricing page.

Seat creep: the bill that grows without a decision

Line four's sibling on team plans is seat creep, and it is worth naming because nobody approves it — it just accumulates. Webflow is the clearest documented case of how this happens: a site plan per site, a workspace plan per workspace, seats on top of the workspace ($39 for a full seat, $15 for a limited one, billed annually — $45 and $19 paid monthly), and add-ons layered over all three — Optimize from $299 a month billed annually ($379 paid monthly), Analyze from $9 ($12 monthly), Localize from $9 ($12 monthly), AI credits from $20 ($25 monthly). Four independent axes, each with its own reason to exist, and a total that nobody set out to reach — it is simply where the sum of reasonable individual decisions lands.

Framer runs a smaller version of the same shape: the base plan is priced per site, and every additional editor is $20 a month on top, so a three-person team working across two sites is paying for two site plans and at least one extra-editor fee before anyone has discussed a budget for either. The fix is not avoiding team plans — it is pricing the team, not the plan, before adding the third or fourth person: multiply the per-seat or per-editor cost by the number of people who will actually need access, not the number on the current invoice, because that number only grows.

Exit cost: what the work looks like on the way out

Line five is the one nobody checks until they need it, which is exactly when it is most expensive to discover. The honest question is not "can I export" but "what comes out, in what shape, and how much of the original work survives the move."

Tool	What you can take with you	What stays behind
reach	The domain you registered, if any	The page itself — no HTML export, no custom code field
Webstudio	Full Remix/React source, self-hostable, AGPL	Nothing — this is the most portable option on this list
WordPress.org	Full site, database and files, GPL software	Nothing structural, though theme and plugin choices affect how clean the move is
Wix	—	Everything; Wix's own support docs say a template cannot be switched on a site already built
Butternut AI	The domain, if bought elsewhere	The site itself — no code export, no WordPress export, no static export of any kind
Lovable	Real React/TypeScript source via two-way GitHub sync, on every plan including Free	Nothing structural, though only one branch syncs at a time

Prices checked August 2026. The spread here is the widest of any line in this worksheet — from Webstudio and Lovable, where the actual source code is yours on every plan, to Butternut, where a site built over months has no export path at all. A subscription price says nothing about which end of that table a tool sits on, and it is the one line item that determines whether switching later is an afternoon or a rebuild from nothing.

A worked example: one small team's year

Put the five lines together for a plausible small team — three people, one shared client-facing site and two personal pages, all real numbers from the two source pages behind this article rather than a hypothetical.

The team runs its marketing site on Webflow: a Premium site plan for the CMS at $25 a month billed annually ($300/year), plus a Workspace Core plan at $19 a month billed annually ($228/year) for the one person who edits it. Two of the three people also want personal pages. One builds theirs on reach at the $49 annual rate, converting a CV into a live page in the time it takes to answer the onboarding form. The other prefers a plain one-page portfolio and picks Carrd's Pro Standard tier at $19 a year, the tier where a custom domain actually becomes available — Carrd's cheaper $9 tier is explicitly for people who don't use one.

Subscriptions alone: $300 + $228 + $49 + $19 = $596 a year. Prices checked August 2026. Add the usage line only where it applies — none of these four plans carries an open-ended credit meter, which is itself worth noticing, since it means this particular team's bill is unusually predictable compared with a team running Bubble or a heavily AI-assisted Webflow workflow. What the $596 does not include, and what no invoice ever will, is the learning curve on the Webflow side — CMS collections, a workspace with roles, a site plan with its own settings — against the near-zero curve on the reach and Carrd side, where there is one page and one owner. That gap is the actual argument for matching tool complexity to job size rather than defaulting to the most capable option available: the cheapest line on this list is not always the smallest cost once the other four are counted.

The exit cost sits underneath all of it, unpriced until the day someone leaves. The Webflow site, if the team ever wants off Webflow, has no clean export path documented anywhere in its pricing pages. Carrd is a one-page site with no build system to speak of, so recreating it elsewhere is closer to retyping than migrating. reach's page would need to be rebuilt from the CV that generated it in the first place, which is a smaller job than it sounds, given that the CV is the actual source of the content — but it is still a rebuild, not an export.

Nothing here argues for the cheapest option in every row. It argues for pricing all five lines before signing anything, because the subscription number is the one line every vendor already put a lot of thought into making look small.