Andrej Karpathy posted the term "vibe coding" in February 2025 — describing what happens when a developer stops reading the code and just rides whatever the model produces. A year and a half later it has its own Wikipedia entry and a multi-billion-dollar product category sitting on top of it. Cursor, Lovable, Replit, and Bolt all sell the same core idea: describe what you want, ship it.
The category collapsed the floor under the freelance MVP gig. It also created the most lucrative new specialty in freelance engineering — taking the apps these tools generate and rebuilding them so they don't fall over in production.
The size of the category in May 2026
The 2025 numbers tell the scale story:
- Cursor crossed $2B in annualized revenue in March 2026, with run-rate doubling in three months and reported funding talks at a $50B valuation according to TechCrunch's coverage of the Bloomberg report.
- Lovable hit $400M ARR by February 2026 with just 146 employees and roughly 200,000 projects built or updated daily, per TechCrunch's Lovable revenue piece.
- Replit ARR was around $253M at the close of 2025 with the Agent assisting 500,000+ unique users in shipping apps, per Sacra's Replit deep dive.
- Bolt.new went from launch to $40M ARR in five months and 5M registered users with roughly 1M daily actives by March 2025, per Sacra's Bolt.new tracker.
Adoption inside the engineering profession itself is the other half of the picture. The 2025 Stack Overflow Developer Survey found 84% of developers using or planning to use AI coding tools — up from 76% the year before. The same survey reports trust in AI accuracy fell 11 points to 29%, and 66% named "AI solutions almost right but not quite" as their top frustration. 45% said debugging AI-generated code is more time-consuming than writing it.
The quality gap is real and getting wider
GitClear analyzed 211 million lines of code across 2020-2024 and found code churn — the share of code rewritten or deleted within two weeks of commit — rose from 5.5% in 2020 to 7.9% in 2024, per their AI Copilot Code Quality 2025 research. Duplicated code blocks grew eight-fold year-over-year. Both metrics are leading indicators of the work needed to keep a codebase maintainable.
Security is worse. At RSAC 2026, Snyk's Manoj Nair reported that AI-generated code creates two to ten times more vulnerabilities per developer, with about 80% of developers admitting they bypass security policies on AI suggestions and only roughly 10% scanning most AI-generated code, per Expert Insights' interview from RSAC 2026.
That gap is the cleanup specialty's market.
What "cleanup" actually means
The work falls into a few stable shapes:
- Architecture rebuild. Founder ships a Lovable or Bolt prototype, gets to a few thousand users, hits a database scaling wall, hires a freelance senior to rewrite the data model and refactor the rendering path before the next traffic spike kills it.
- Auth and permissions hardening. AI builders ship working auth and a "users" table. They do not ship multi-tenant isolation, RBAC, audit trails, or session-revocation flows. Founders selling to anyone with a procurement department need this work.
- Payments and billing remediation. Stripe in test mode plus a hardcoded webhook handler that doesn't verify signatures is the typical baseline. Going live is a remediation project.
- Performance and observability. Logging, error tracking, query plan inspection, caching, rate limiting — all the things the AI builder did not generate because the prompt didn't ask for them.
- Compliance overlays. GDPR data export and erasure, US state privacy law (CCPA, CPRA), HIPAA / SOC 2 prep work. The AI builder did not generate a data inventory.
The rate card in May 2026
Median rates for the cleanup specialty are tracked across Toptal, Upwork, and the senior-only marketplaces. Aggregator data on Gigradar's Upwork hourly rate report places senior AI engineers at $100-$200/hr, AI agent development at $175-$300/hr, and RAG implementation at $150-$250/hr. Toptal's senior tier publishes a similar range — the cleanup work sits inside it because the skill set is the same.
The pricing pattern that holds in 2026:
- Discovery + audit, fixed-fee. $2,500 to $5,000 for a one-week assessment of a Lovable / Bolt / Replit codebase, ending in a written remediation plan.
- Hourly cleanup, $150-$250/hr. Most of the actual work bills here. The freelancer is being paid to think through architectural choices the AI builder did not.
- Project-priced rebuilds, $30K-$80K. A 4-8 week engagement to rewrite a working but fragile prototype as a maintainable production system. The senior tier was the floor for this work pre-AI; it is now the typical engagement.
The market dynamic is the same one we covered in our piece on Lovable, v0, and Bolt eating the $5K MVP market: the floor compresses fast, the senior tier expands faster.
Why this category is durable
The Anthropic Economic Index for March 2026 reports that 36% of all Claude.ai usage is coding, with Computer & Mathematical occupations contributing 35% of conversations and 44% of API traffic, per Anthropic's index report. Coding is the workload AI does most of, and the workload that produces the most code that needs to be checked.
The cleanup specialty isn't a transitional opportunity. It's structural:
- Volume of generated code keeps growing. Cursor and the AI builders ship millions of new codebases per month.
- Each codebase has a non-zero probability of going to production.
- A non-trivial slice of those production codebases will outgrow what the AI builder shipped.
- The senior engineers who can rebuild them are scarce.
The closest historical analog is the late-2000s WordPress cleanup boom — every small business had a WordPress site somebody else built, and a generation of freelancers made a living rebuilding them.
How to position for cleanup work
The pattern that works for solo senior engineers in this niche:
- Lead with a fixed-fee audit. Founders shopping for cleanup work are skittish about open-ended hourly bills. A $3,500 one-week audit with a written remediation plan is an easier first sale than "$200/hr, terms TBD."
- Show the rebuild, not the prototype. Case studies that walk through a Lovable / Cursor / Bolt prototype's specific failure modes and the fixes, with code excerpts, convert much better than generic "senior engineer for hire" pitches.
- Pick a vertical. Healthcare, fintech, B2B SaaS — pick one. The cleanup work is similar across verticals at the architecture level, but the compliance overlays and the buyer's vocabulary are not.
- Charge for the rebuild plan, not the rebuild. A documented architecture for a multi-tenant rewrite is what founders pay $3K-$5K for; the rebuild itself prices off that document.
Delivvo gives senior freelance engineers a single branded portal for the proposal, fixed-fee audit contract, file delivery, and invoicing on the cleanup engagement that follows — so when the founder is paying $200/hr to rebuild the prototype their AI builder shipped, the engagement around the rebuild looks like a senior engineering firm and not an Upwork escrow. See how it works →
The shorter version
Vibe coding generated a category. The category is generating apps faster than the senior engineering tier can rebuild them. That gap is the highest-paying freelance specialty of 2026 — not because the rate cards changed, but because the volume of work pricing into the senior tier is the largest it has ever been.
Written by The Delivvo team · May 10, 2026
More from the blog →
