The total at the bottom of the Replit invoice read $480.27, and my Pro plan was supposed to cover the first $100.

That gap is the story. I gave Replit Agent 3 seven specific startup ideas — the kind of thing I’d ordinarily hand to a junior contractor and a Linear board — and asked it to ship each one as an actual deployed app. Eight days. One credit card. No babysitting beyond clicking “yes, continue.”

Four shipped. Two were unusable. One I killed mid-build because the meter scared me. This is the report, not a benchmark.

The setup

I bought Pro annual at $95/mo. That comes with $100 in monthly credits. Replit switched to a credit + effort-based pricing model back in mid-2025, and the model is a touch opaque — verified bills tend to show fixed rates per checkpoint, not the truly metered “you pay for what the agent thinks about” thing the marketing implied. One Reddit thread I read had a user reporting the same app idea jumping from $0.50 to $3 on the first prompt after the change. I went in expecting variance.

Seven ideas, all my own backlog. Not Replit’s gallery, not Twitter trends. Each one I’d actually consider shipping. I wrote each brief in plain English, around 80–120 words. I let Agent 3 pick its own stack. I refused to intervene on file structure or library choice unless a build broke for 3+ retries.

The seven, scored

Total time: 19h 37m. Total spend: $480.27. Plan covered $100. Overage: $380.27 — billed at the end of week one, which is when I noticed the credit had vaporized on idea #2 alone.

The four that shipped — what they’re like

1. RSS-to-Slack summarizer

The platonic ideal of an Agent 3 task. One Python script, one Slack webhook, one Replit cron. Agent picked FastAPI, which was overkill, but it ran. I have it summarizing five SaaS-pricing blogs daily into a channel called #price-watch. It’s been live 11 days, zero failures.

3. Twitch overlay generator

Surprised me. I’d assumed Agent would flail on canvas/WebGL stuff. It picked PixiJS, wired Twitch OAuth, and produced a functional overlay configurator in under two hours. My friend who streams Valorant has been using it for a week. He paid me $5. I am now a profitable Replit Agent reseller, technically.

4. Whisper + Claude meeting notes

Worked. The auth flow was a mess — Agent built Supabase auth, then halfway through decided to switch to Clerk, then half-undid it. I had to manually delete the dangling Clerk components. Final product transcribes a 30-min meeting in about 90 seconds and gives back a clean action-item list. I’d use it. Don’t know if I’d pay for it.

6. Regex explainer extension

Honestly impressive. Agent generated the manifest, the popup, the content script, and even a usable icon. I submitted it to the Chrome Web Store the same evening. It cleared review in 18 hours. It’s not making money. It’s making 14 daily users smile, which is something.

The three that didn’t

2. Dog-walker marketplace (the $118 lesson)

Agent built the UI, the seeker dashboard, the walker onboarding, the search. All decent. The moment I asked it to wire Stripe Connect for marketplace payouts, it spent 38 minutes generating, rewriting, regenerating four versions of the webhook handler, each one almost correct, none of them actually able to onboard a connected account. The bill ran while it deliberated.

This is the “effort-based” pricing in practice. The harder the agent thinks, the more you pay, whether or not the thinking lands. I should have stopped it at the 20-minute mark.

5. Receipt manager (iOS Shortcut + web)

I killed this one. The brief required Agent to scaffold a web companion that knew about an iOS Shortcut it could not see. The agent kept asking clarifying questions, then trying to design around them. Each clarification cost. I burned $89.60 to learn that the agent doesn’t really do cross-platform spec inference well.

7. Agency CRM (crashed three times)

The most expensive failure. Agent built a perfectly fine Next.js + Prisma CRM with roles, deal pipelines, and an inbox. It would not deploy. Migrations failed on the production Postgres each time. The agent’s autonomous debugging — much hyped — got into a loop where it would “fix” the migration by adding a column the schema didn’t expect, then “fix” the new error by removing a column it now needed elsewhere. Three full rebuild cycles. $135.77.

I gave up and shipped it to a Railway box manually. Took me 22 minutes. So the Agent did 95% of the work and then immolated itself on the last 5%.

Cost vs the alternative

The contractor math is rough — $60/hr for a competent junior, 40 hours of work, gets you maybe one more shipped app than Agent 3 and a meaningful amount of human follow-through. So Agent is cheaper. It is also worse at the hard stuff and identical at the easy stuff. The ratio matters less than the friction: Agent ships at 3am on a Saturday and a junior does not.

What Agent 3 is and isn’t, plainly

• Is: A very good first-draft author of single-system, well-scoped apps. Cron jobs, dashboards, Chrome extensions, internal tools, anything where you can write the spec in a paragraph.
• Isn’t: A reliable shipper of anything involving cross-system contracts (payouts, OAuth dances with multiple providers, migrations against live data). It will try, you will pay for it to try, and the result will be a confident-looking thing that fails its first test.
• Doesn’t tell you: When it’s spending unproductive effort. I’d pay double for a “this is going badly, stop me” guardrail. The verified-bill thread on Reddit suggests checkpoints are charged at fixed rates rather than truly effort-based, which makes the marketing positioning slightly misleading.

Reviewers recommend Replit without hesitation for rapid prototyping, learning to code with AI assistance, and collaborative development. However, pricing escalates quickly for professional use, and Agent 3 sometimes requires manual intervention on edge cases.

Hackceleration’s 2026 Replit review — matches my experience exactly

The blind spot I owe you

I didn’t test the High Power Model toggle that adds extended thinking. Two reasons: it was already gobbling credits faster than I could account for, and Replit’s docs were unclear at the time I ran the test whether HPM was on by default for Pro. If it was on, my $480 was actually cheap. If it was off, my $480 should have been closer to $240 and I’m being gouged by my own checkbox. I’ll redo idea #2 next month with HPM explicitly toggled, on a fresh account, and post the gap.

What happens next

Replit is in a strange spot. Agent 3 is unambiguously good at one category — single-system app generation — and unambiguously not yet good at the next category — multi-system orchestration. The pricing model bets on the second category coming. If they ship a v4 that actually handles Stripe Connect, OAuth chains, and migrations in a single brief, the $480 stops looking ridiculous and starts looking cheap. If they don’t, the $480 looks like a tax for being early. My bet: v4 lands by AI Engineer SF in June with explicit pricing for the multi-system case, and quietly walks back the “effort-based” framing to a clearer per-action one. If the new framing makes my idea #7 cost $40 instead of $135, I come back. If not, I’m spending the $95/mo on a Hetzner box and Claude Code, which gave me four shipped apps last weekend for $11 in tokens.

That comparison is going to be the next test.