AnalysisNew

    Your App's AI Feature Runs on Four Credits and a Key You Don't Own

    In four days in September, Lovable put Anthropic's Claude models and two new GPT-6 models inside the apps it builds. The feature is real and it is well engineered. The monthly grant behind it is four credits, and the failure mode it hands your users is an HTTP status code.

    An AI feature used to be the most expensive thing on a prototype's roadmap. You needed a provider account, a billing relationship, a key-management story, and somewhere server-side to put the call. Over three days in September 2026, Lovable removed all four of those obstacles. What it did not remove is the operating question underneath them, and that question now arrives later, in front of users, wearing an HTTP status code.

    What landed, 23 to 25 September

    On 23 September, Lovable's changelog records that your app's AI features can now use Anthropic's Claude models, listing Claude Fable 5.1, Claude Opus 5.5, Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5. The same day, image generation and video generation began appearing in Usage details as separate line items under Build credits. On 25 September two OpenAI models followed for the same purpose, GPT-6 Sol and GPT-6 Luna, both of which accept text and image input and respond with text. In between, on 24 September, the members export gained Effective Credit Limit and Credit Limit Source columns, and the Manage instance dialog began showing an estimated monthly Cloud cost in credits under each database size.

    Read those four entries together and a theme appears that is not about models at all. Three of the four are about credits.

    The engineering here is good, and that matters

    It is worth being precise about what Lovable built, because the criticism that follows is narrow and the thing it is aimed at is otherwise solid.

    Lovable's documentation states that it "automatically generates and manages a LOVABLE_API_KEY for each project" and that "You never need to create or provide it yourself." The calls do not run in the browser: "AI calls run through a secure backend edge function that Lovable creates for you. Calls are not made directly from the browser, which keeps your credentials and prompts server-side." And the provider relationship is gone entirely, with the documentation stating that "You do not need to create a provider account, configure billing with a model provider, or paste API keys into your app."

    Each of those closes a class of mistake we see in prototypes: the API key in the client bundle, the key committed to the repository, the provider account attached to a founder's personal card. Lovable has made all three structurally impossible on this path. That is a real improvement and it should be said plainly.

    The number

    Now the number. Lovable documents a 4-credit monthly AI grant for AI gateway usage in deployed apps, on Free, Pro and Business workspaces. Past the grant, AI gateway usage draws from general credits where available. And when there are no credits left, the documentation is specific about what happens: the gateway returns 402 Payment Required.

    That last sentence is the whole post. Most limits in a prototype degrade. A rate limit slows you down. A cache miss costs a hundred milliseconds. This one does not degrade. It is a hard stop, drawn against a shared budget that is also being consumed by the image generation, the video generation and the general build work now itemised alongside it, and it surfaces at the one moment nobody is watching: after launch, in production, to somebody who is not you.

    Two ways to run the same feature

    QuestionLovable AI gatewayYour own provider account
    Who holds the credentialLovable, as a per-project LOVABLE_API_KEY it generates and managesYou
    Where the call runsA backend edge function Lovable creates, not the browserWherever you put it, which is now your decision
    Setup before the first callNone documented: no provider account, no billing setup, no pasted keysAccount, billing, key storage, server route
    Monthly floorA 4-credit AI grant on Free, Pro and Business, then general creditsWhatever your provider bills
    When the budget is gone402 Payment Required from the gatewayYour provider's own error, on a budget you set
    Who can raise the ceilingWhoever controls the workspace's creditsWhoever controls the billing account
    What happens on code exportNot stated on Lovable's AI features page; confirm before you depend on itNothing, the account is yours

    Four questions to answer before this ships

    This is not an argument against the gateway. For most prototypes it is the right call, and the alternative reintroduces exactly the key-handling problems it removed. It is an argument for answering four questions before the app is in front of paying users.

    1. What does the app render when the call returns 402?

    Not what it logs, what a user sees. A spinner that never resolves is the default outcome and it is the worst one. The honest version is a short message saying the feature is temporarily unavailable, on a path that does not discard the user's input.

    2. Who looks at the credit balance, and on which day?

    The grant is monthly. A monthly budget with no owner and no alarm is a budget that gets discovered empty. Lovable added Effective Credit Limit and Credit Limit Source to the members export on 24 September, which makes the number exportable. Exportable is not monitored.

    3. Can a logged-out visitor reach the AI call?

    If the answer is yes, consumption is no longer bounded by your user count. It is bounded by your traffic, including the traffic you did not ask for. This is the check we would run first.

    4. Did anyone choose the model?

    Five Claude models and two GPT-6 models arrived within three days. A prototype prompted into existence before 23 September is running whatever the default was then. That may still be the right model. It should be a decision with a reason attached, because the model is also a line on the credit draw.

    Why this lands on prototypes

    The platform keeps adding visibility: the cost estimate under each database size, the separate line items for image and video, the credit limit columns in the export. That visibility is genuinely useful and Lovable is shipping it quickly. But visibility is not a budget, it is not an alert, and it is not a fallback path in your interface. Those three are application work, and they belong to whoever owns the app.

    Lovable is also not alone in re-cutting its units this month. On 25 September Supabase introduced usage-based pricing for log ingest and query volume, with a grace period into early 2027. On 23 September Vercel made Blob stores unlimited on every plan and made creating one cost an Advanced Operation instead. Neither is a price rise in the ordinary sense. They are the same move: something that used to be free-shaped becomes a metered unit, and an app built before the meter existed carries on as though it still does not.

    An AI feature is the newest thing a prototype can contain and the first thing a user notices when it stops. It deserves what the payment path gets: a known failure mode, a person watching the meter, and a behaviour you chose rather than inherited.

    Getting these four answered

    IdeaFlo has completed 80 jobs on Upwork, turning Lovable and Vite prototypes into applications that hold up in production. If you want the four questions above answered against your project, alongside the rest of the pre-launch inventory, the Strategy Audit is $499 one-time and is credited toward any build. If the app needs to become a real application first, the Publishable MVP is $4,999 one-time, and the Founding Client package is $2,999 one-time with a 14-day build guarantee and a full refund if the deadline is missed. Every build carries the same promise: first paying customers within 90 days of launch, or a full refund.

    If we are not the right fit, we will say so on the call.

    Book your free 20-minute intro call

    Sources

    Related posts