Every vibe coding tool prices the same illusion: that the cost of an app is the cost of generating it. The first prompt is cheap and spectacular. The budget dies later, in the fix loop, and the fix loop is not an edge case. It’s the normal operating mode of every code-generating tool on this site once an app passes trivial size.
Why the 20th prompt costs more than the 1st
The first prompt writes onto a blank canvas; the model is good at that, and one round usually buys visible progress. The 20th prompt has to modify an existing system the model only partially remembers. As codebases grow they exceed the AI’s working context, and the model starts contradicting its own earlier decisions. Fixes address symptoms rather than root causes, so a patch in one place breaks another, the pattern builders call prompt whack-a-mole. Iteration also bloats the artifact itself: limited context means the AI rewrites utility functions it can’t see, leaving duplicate logic and a patchwork of styles that makes every subsequent fix harder to land.
So the unit economics invert. Early prompts buy features; late prompts buy attempts. And every attempt is billed.
What the meters say
The research numbers, by tool, all from documented user reports.
Lovable sells credits, base Pro at 25 euros for 100 a month. Users report per-prompt consumption rising from around 1.2 credits to 3-4, a roughly tenfold cost inflation over time, with even questions about the code consuming fractions of a credit. Reviewers describe the canonical loop: credits spent on debugging chats where the agent introduces new errors while resolving the first, and reports of it claiming a fix happened when it didn’t. At 3-4 credits per prompt, a 100-credit month is under 30 attempts.
Bolt sells tokens, 10 million on the $25 Pro tier. Its signature complaint is paying for non-progress: the diff-edit that gets immediately rewritten without the change, “just burning tokens with no changes,” and a monthly limit consumed by a generated error, leaving the builder waiting for next month to fix the tool’s own mistake. Reviewers also describe opaque depletion during complex loops, with no breakdown of which edits ate the tokens.
Replit is the steepest curve because it’s effort-priced: the bill tracks how hard the agent works, and nothing makes an agent work harder than debugging itself. Documented cases: $25 of credits in under a day, $350 in a single day, $700 in a month, and $1,500 in surprise database charges driven partly by per-checkpoint backups. The community’s darkest reading is structural: more mistakes mean more fixes mean more billable runs.
Different meters, same shape. The pricing unit is the attempt, and debugging is the activity that maximizes attempts.
The tax, named
Call it the fix loop tax: the gap between what you’d pay if generation worked the first time and what you actually pay. It has three properties worth writing down. It’s invisible at purchase time, since the sticker price describes the happy path. It’s regressive, hitting hardest the builders least able to diagnose root causes, because they run the most rounds. And it’s correlated with importance: the apps that trigger the deepest loops are the auth-heavy, edge-case-rich business apps you most need to work, the kind examined in Lovable vs Bolt.
The tax also doesn’t stay in the cost column. Each billed round on an auth-adjacent feature re-rolls the security dice from what ‘45% of AI code is vulnerable’ actually means, and the loop itself is the day-to-day shape of the Day Two problem: the maintenance phase where every change risks breaking the last one. The meter is just the part you can see on the invoice.
Paying less, or not at all
On code-gen tools the mitigations are craft: scope prompts narrowly, commit every working state, read diffs before accepting, and recognize drift early enough to stop. A flat-subscription tool like Cursor at least bounds the worst case at a known monthly number.
The structural answer is to notice which apps don’t need a loop. A portal or internal tool is mostly auth, permissions, and CRUD, and on a platform like Softr those are configuration: change a setting, the change is made, no regeneration round and no attempt billed. Softr has AI credits for its Co-Builder, but since everything the AI does can also be done manually, an empty balance never blocks a fix. For business apps, the cheapest fix loop is the one that doesn’t exist.