The bill nobody shows: vibe-coding credits, tokens, and markup (Replit, Bolt, Lovable, v0)

The vibe-coding bill never arrives as a plan price. It arrives as a sum of meters spinning per prompt, per fix, per reloaded context. Whoever watches only the entry price understands little, because the monthly plan is only the front door and real consumption happens in use. The reports gathered here show the pattern with hard numbers, from $1k Replit weeks to 20M tokens for one Bolt auth bug. For the month's full view, with operation and rework, start with the what vibe coding really costs guide. This piece opens the black box of all four billing rulers and shows where each bites, with values, dates, and sources so you can check every line.

Replit shows the group's most direct markup. The Core plan costs $25, Agent hours cost $30, and reports describe weeks near $1k against $180-200 before Agent 3, launched Sep 2025, with one stray $114 overage (The Register via brgrowthclub, Mar 13, 2026). The plan never vanished. It stopped being a ceiling and turned into a floor with the meter running. The $180-to-$1k gap across seven days is no wrong click. It is a regime where every attempt, every fix, and every reloaded context adds hours plus overage.

Hourly math explains the blowout speed. At $30 an hour, $1k in a week equals 30+ billed hours in seven days, an easy pace once the Agent runs, misses, retries, and reloads the whole project every lap. The $114 overage shows the same regime's second blade. Even clock-watchers pay overage once usage passes the plan bound. The meter never separates good attempts from useless ones. It logs time plus consumption, period.

The pattern repeats outside Replit. There is Fred Benenson's report past $1k on an AI-generated project (Fred Benenson). With no round week-or-month date in the cited report, the value matters for what it signals. Four digits to lift a project off paper with AI is no isolated exception. It portraits a flow where building looks cheap on day one and bills heavily across rounds. Where the ruler is time plus overage, insisting on one mistake costs more than restarting on a smaller scope.

Bolt bills on context, never clicks. The surveys cite 10K tokens per prompt on fresh projects against 100K once the project reaches 20 components, with routines near 1.3M tokens a day (brgrowthclub, Mar 13, 2026). The curve outrages because it runs silent. Ten times the tokens for the same tweak request, only because the app grew. Every fresh component joins the context, and every fresh attempt pays freight on the whole app again. The prompt never grew 10x. The freight did.

Mistakes show where that freight detonates. The same surveys cite 7-12M tokens for plain mistakes and 20M tokens for one auth bug (brgrowthclub, Mar 13, 2026). A plain mistake burning up to 12M tokens runs anything but plain on the wallet. One login bug burning 20M tokens equals 2,000 fresh 10K prompts, or 200 full 100K prompts. Auth hurts most because sessions, permissions, redirects, and outside providers tangle, parts the model guesses at instead of reading as contract. Every wrong guess reloads everything and retries in new wording.

There is also the ghost deploy tax. Reports cite ghost files trailing into Netlify, generated files serving no app purpose yet entering consumption plus deploy (brgrowthclub, Mar 13, 2026). Dead code never sits neutral. It occupies context in later rounds, confuses the model about what runs, and bills twice, once to generate, once to load in later attempts. In Bolt, cleaning the project is no organizing whim. It directly cuts tokens per prompt.

Lovable looks most predictable in the window. The Pro plan ships 100 credits for $25, and surveys add Supabase at $25 plus $10 per app (Softr via brgrowthclub, Mar 13, 2026). The one-app base already touches $60 a month before any blowout, $25 builder plus $25 database plus $10 app. Every extra app adds $10. Credits cover builder chat. Data lives on another subscription, another ruler. Comparing only the Pro price compares half the bill.

The second biter is rework. The same surveys point 30-40% of usage into bugfix (Softr via brgrowthclub, Mar 13, 2026). Per 10 credits, 3 to 4 pay fixes, never creation. On a 100-credit plan, 30 to 40 credits repair what the generator shipped crooked. The ratio outrages because it flips the promise. The tool sells creation speed and bills a third of consumption repairing the delivery. Credits never measure only output. They measure defects too.

The working read sums both layers before subscribing. A fixed base near $60 with one app, plus variable burn with nearly 4 in 10 credits cycling back to one spot. Loose scopes worsen both ends at once. More screens demand more database plus rules, lifting the base, and breed more bug surface, lifting the fix share. Locking scope plus splitting auth and data out of screen prompts cuts both bills at once.

v0 runs the most technical ruler, hence the easiest to underestimate. The cited May 2025 values are $7.50 per million input tokens and $37.50 per million output, with hidden context you never see in the prompt (brgrowthclub, Mar 13, 2026). Millions sound distant until they turn routine. At 1.3M a day, the intense-use level in the same survey set, one day exceeds a billed million. Small units hide fast sums.

The in-out asymmetry outweighs the absolute price. Output costs 5x input, $37.50 against $7.50 per million. Long, diffuse answers cost far more than long, precise questions. When the model returns whole files to adjust a few lines, rewrites working components, or generates long explanations beside code, the bill runs in the top band. Asking less with scoped bounds is no etiquette. It is direct arbitrage across both bands.

Hidden context closes the trap. Part of the bill never shows in the text box, like system instructions, history, and files the flow itself attaches (brgrowthclub, Mar 13, 2026). Users control what they type, never everything billed. So honest per-token pricing comparisons must include the invisible. Two text-identical prompts can bill differently as history and attachments shift. Budgeting by what you wrote undercounts by ruler design.

The most aggressive number in the cited tests is no price table. It is 211k tokens burned across 3 fixes with nothing repaired, against 39k tokens at initial build, a 5.4x multiple (TabNews, Feb 26, 2026). The build cost a fraction. Stubbornness cost the multiple. The cited test also logs the detail explaining repetition. The app looked nearly done, every tweak promised to close, every fresh attempt returned the defect in new wording. Insisting down one thread with poisoned context turned a small mistake into a big bill.

the 5.4x ties all four rulers into one image. Hours, context, credits, and millions bill differently, and all bill repetition. Whoever stops the third unrepaired attempt cuts the markup at the root, on any tool.

The chart below scales that disproportion. The short bar is the cost of creating. The long bar is the cost of insisting with no strategy change.

The defense is procedural and holds across all four rulers. Freezing scope once yield stalls, opening fresh threads with minimal cuts instead of dragging poisoned history, reviewing outside the generator before publishing break the documented cycle. The third unrepaired attempt is the stop sign, never the fourth invitation. Whoever switches strategy at the second miss pays a fraction. Whoever repeats one request pays the multiple, in Replit hours, Bolt context, Lovable credits, v0 millions.

The table below compresses all four rulers with values plus the main trap. Each row answers one question in a distinct currency. Where the meter spins fastest once the project stalls.

Replit, Bolt, Lovable, and v0 values via brgrowthclub, Mar 13, 2026, with v0 values from May 2025. The 211k-versus-39k loop via TabNews, Feb 26, 2026. The past-$1k case via Fred Benenson.

The rule against paying markup fits four locks. First, round caps per chore, two attempts down one thread, the third in a fresh thread with a minimal cut. Second, auth and data split out of screen prompts, because login plus database are the spots turning 20M on Bolt and 30-40% bugfix on Lovable. Third, dead-file cleanup before fresh rounds, never paying freight on useless code nor exporting trash to Netlify. Fourth, reviews outside the generator before publishing, with a login, permission, and billing checklist.

To see how those rulers behave on one challenge, with cost per delivery instead of list prices, read Lovable vs Bolt vs v0 on the same app. Per-attempt math deceives. Per-delivery math decides.

What does vibe coding really cost per month?

It follows the ruler plus the loop, never the plan alone. Reports cite Replit Core at $25 with Agent hours at $30 and $1k weeks against $180-200, Lovable Pro at $25 plus Supabase at $25 with $10 per app, Bolt at 1.3M tokens daily in intense use, v0 at $7.50 and $37.50 per million (brgrowthclub, Mar 13, 2026). Add rework to the base, 30-40% on bugfix plus cases of 211k against 39k (TabNews, Feb 26, 2026).

How do Lovable credits work?

The Pro plan ships 100 credits for $25 to chat with the builder, and data sits on a separate Supabase subscription at $25 plus $10 per app (Softr via brgrowthclub, Mar 13, 2026). In practice the one-app base touches $60 a month. Of that burn, 30-40% goes to bugfix, so 30 to 40 credits per 100 pay fixes instead of creation.

What is the doom loop and why does it weigh most on Bolt?

Doom loop is the fix sequence repairing nothing while still billing. The cited test measured 211k tokens across 3 unrepaired fixes against 39k at initial build, or 5.4x (TabNews, Feb 26, 2026). On Bolt the weight runs heavier because every attempt reloads the entire context, 10K on fresh projects to 100K with 20 components, turning plain mistakes into 7-12M and auth bugs into 20M (brgrowthclub, Mar 13, 2026).

How does v0 per-token pricing work?

v0 bills per million tokens, $7.50 in and $37.50 out at May 2025 values, plus hidden context outside the prompt (brgrowthclub, Mar 13, 2026). Output costs 5x input, so long answers and broad rewrites weigh most. Since part of the billed context never shows in the text box, two text-identical prompts can bill differently as history and attachments shift.

The vibe-coding bill never runs dear by table design. It runs dear through the repetition each table bills its own way. Hours on Replit, context on Bolt, credits with an attached subscription on Lovable, millions with hidden context on v0. The numbers size the trouble, $1k weeks and 20M on one auth bug through 30-40% on bugfix and 5.4x between insisting and building. The lock matches across all four rulers. Short scopes, separate auth, fresh threads after two misses, reviews outside the generator before publishing.

professional flow from prototype to deploy

• The Register via brgrowthclub, Mar 13, 2026: Replit Core $25, Agent 3 from Sep 2025, $1k weeks vs $180-200, $30 hourly, $114 overages.
• Softr via brgrowthclub, Mar 13, 2026: Lovable Pro 100 credits $25, Supabase $25 plus $10 per app, 30-40% of usage on bugfix.
• brgrowthclub, Mar 13, 2026: Bolt 10K per fresh prompt vs 100K with 20 components, 1.3M daily, 7-12M plain mistakes, 20M auth, ghost files on Netlify; v0 May 2025 $7.50 in and $37.50 out per million, plus hidden context.
• Fred Benenson: past-$1k spend on an AI-generated project.
• TabNews, zilvodev, Feb 26, 2026: 211k tokens across 3 unrepaired fixes vs 39k at start, 5.4x, doom loop.