afterbuild/ops
ERR-915/stack trace
ERR-915
Bolt.new Burning Through Millions of Tokens on One Bug? How to Stop the Spiral

Bolt.new Burning Through Millions of Tokens on One Bug? How to Stop the Spiral

Last updated 15 April 2026 · 9 min read · By Hyder Shah
Direct answer

Your Bolt.new project spirals the moment the AI can’t see your whole codebase. Public reports from other users include “20 million tokens spent trying to fix a single authentication issue” and “$1,000+ on a single project”. You stop it by freezing your code, reproducing your bug outside Bolt, fixing it by hand, and using Bolt only for UI iteration. Five concrete steps below, each under 30 minutes.

Quick fix for Bolt.new Burning Through Millions of Tokens

Start here

Freeze the code — commit the last known-good state

In Bolt, push your repo to GitHub immediately. Tag it last-known-good. If your app currently works on the happy path, that tag is your rollback point. Every future regression is now cheap to undo.

This single step stops the most expensive failure mode: prompting the AI to fix something and silently losing yesterday’s working feature.

Deeper fixes when the quick fix fails

  1. 02

    Pull the code out of Bolt and reproduce the bug locally

    Download the zip, or git clone the repo you just pushed. Run npm install and npm run dev. Bolt is a WebContainer; your laptop is real Node. If the bug disappears locally, it’s a WebContainer quirk, not a real bug — skip to Step 4 and deploy from local.

    If the bug reproduces locally, you now have a real debugger, real logs, and zero token cost per iteration.

  2. 03

    Fix the one root cause by hand, then write one test

    Whatever the bug is, diagnose with browser DevTools and VS Code, not by prompting. Fix the single root cause. Then, critically: write one automated test that fails without your fix and passes with it. Vitest or Playwright, one file, one assertion.

    This test is the signal Bolt was missing. Now when you go back to AI-driven iteration, you have a canary for this exact regression.

  3. 04

    Deploy from GitHub, not from Bolt

    Connect your GitHub repo to Netlify or Vercel. Make the deploy trigger on push to main. Bolt now becomes an optional editing surface, not the single point of failure for your production site.

    This kills the “every new deployment deploys into another universe”failure mode — each deploy is now a git commit you can inspect and roll back.

  4. 05

    Constrain Bolt to UI-only iterations

    Use Bolt for what it’s actually good at: visual scaffolding and component iteration. For anything that touches auth, payments, webhooks, or your database schema, edit locally in VS Code with Claude Code or Cursor.

    Our internal rule: if the change touches a file under /api, /lib/auth, or any .sql migration, do it by hand. The token cost of Bolt solving those correctly is higher than the hourly rate of a senior engineer.

If you’ve already spent more than $500 on a single bug, the math is simple: a fixed-price fix from a human engineer is almost always cheaper than one more week of token spend.

Why AI-built apps hit Bolt.new Burning Through Millions of Tokens

Every prompt you send in Bolt reloads your entire project into the model’s context. As your app grows past a few hundred lines, each “fix this” prompt you send costs more tokens and the model’s attention on your code dilutes. You land in the fix-break-fix loop: resolving your error A creates error B, resolving B reintroduces A. Each cycle costs you more than the last.

One Trustpilot reviewer captured what you’re seeing exactly: “It feels like a slot machine where you’re not sure what an action will cost.”Your root cause is architectural, not accidental — the model simply cannot hold enough of your real app in its head to debug it reliably past a certain size threshold.

20 million tokens spent trying to fix a single authentication issue. Every time, I just throw my money away.
Bolt.new user, Trustpilot

Diagnose Bolt.new Burning Through Millions of Tokens by failure mode

Match your current symptom to the row below before touching the prompt again. Every row maps to a step you’ll run below.

SymptomWhat's really happeningStep
Fixing error A creates error B, then B brings A backRegression loop; context window full of stale decisionsStep 1
Last 3 prompts cost 200k+ tokens eachContext is now dominated by unrelated filesStep 2
AI 'fixed' the bug but on redeploy it's still brokenFix was only in preview memory, not committedStep 3
You've spent 5M+ tokens and feature still incompleteProblem is outside Bolt's reliability zoneStep 4
Every 'improve this' prompt degrades working codeNo tests; AI has no signal for what must not changeStep 5

Related errors we fix

Still stuck with Bolt.new Burning Through Millions of Tokens?

Emergency triage · $299 · 48h turnaround
We restore service and write the root-cause report.

If the math is no longer working, a fixed-price fix stops the bleed:

  • You've spent >$500 on one bug
  • You've burned >5M tokens on a single feature
  • You're in a visible fix-break-fix loop
  • You're scared to prompt again because of regressions
start the triage →

Bolt.new Burning Through Millions of Tokens questions

Why does Bolt.new burn through so many tokens on one bug?+
Every Bolt prompt reloads the entire project into the model's context. As the app grows, each prompt costs more and the model's attention dilutes. You get the fix-break-fix loop where resolving error A creates error B, and B reintroduces A. Real user reports include 20 million tokens on a single authentication issue and $1,000+ on a single project.
How do I stop Bolt.new from breaking working features?+
Three moves: (1) push to GitHub and tag the last-known-good commit before any major prompt, (2) add at least one automated test so regressions are caught, (3) constrain Bolt to UI-only iterations and do auth, payments, and database work by hand in VS Code. These three steps turn Bolt from a slot machine into a tool.
Is Bolt.new worth the token cost?+
For UI scaffolding and small prototypes under 500 lines, yes. Past that point, the $1,000+ reports start appearing in reviews. The sweet spot is using Bolt to sketch the first draft, then exporting to GitHub and finishing locally. Users who stay inside Bolt for the full lifecycle almost always regret the bill.
Can I get a refund for wasted Bolt.new tokens?+
Bolt's refund policy is discretionary. Users on Trustpilot report mixed success. The reliable path is prevention: freeze the code at every working state, deploy from GitHub, and escape Bolt the moment you detect a regression loop. Past a certain spend, the money is gone — focus on stopping the bleed.
How much does it cost to rescue a token-drained Bolt.new project?+
Fixed-price rescue starts at $299 for an Emergency Triage (48h, single root cause identified), $3,999 for Break-the-Fix-Loop (we escape Bolt, ship the feature, add tests, hand you back a stable repo), or $7,499 to finish a full MVP. Most token-drained projects take 8-30 engineer hours to close out.
What AI tools waste fewer tokens than Bolt.new for debugging?+
For debugging specifically, Claude Code or Cursor with a scoped context window outperform Bolt on cost per fix, because the developer controls which files load. Bolt's token math assumes whole-project reload every prompt, which is wasteful for small fixes. Use Bolt for greenfield UI, Cursor/Claude Code for bug work.
Next step

Ship the fix. Keep the fix.

Emergency Triage restores service in 48 hours. Break the Fix Loop rebuilds CI so this error cannot ship again.

About the author

Hyder Shah leads Afterbuild Labs, shipping production rescues for apps built in Lovable, Bolt.new, Cursor, Replit, v0, and Base44. our rescue methodology.

Bolt.new Burning Through Millions of Tokens experts

If this problem keeps coming back, you probably need ongoing expertise in the underlying stack.

Sources