Replit Agent 3 Review: The App Builder That Ships in an Hour and Bills Like It Knows

What it’s actually good at

The core loop is what makes Agent 3 feel like a genuine leap. You type a plain-English description of an app, hit send, and watch it scaffold. On the waitlist form, I gave it three sentences (capture email and name, gate it behind a Stripe pre-order, send a confirmation email) and about 34 minutes later I had a working, deployed app with a Neon Postgres database behind it and a Stripe checkout that actually cleared a test card. That’s not a demo. That’s the thing shipping.

The integration story is the piece I didn’t expect to love. Agent 3 spawns sub-agents for workflow automation and supports 160+ third-party integrations via OpenInt, which Replit acquired in October 2025, covering Stripe, Twilio, Google Sheets, Slack, and over 150 others, all connectable from a single prompt without writing any authentication code manually. In practice, that means you say “wire this form to my Google Sheet” and it does. No OAuth dance, no fumbling with a Zapier hop.

Autonomous runtime is the other quiet upgrade. Agent 3 launched in September 2025 and extended autonomous runtime tenfold.

Given a prompt, it creates files, installs dependencies, sets up a Neon Postgres database, writes authentication logic, and deploys, operating autonomously for up to 200 minutes per task. The old Agent 2 would tap out on anything ambitious and hand you a half-built skeleton; Agent 3 will actually push through a multi-step build in one session, run its own tests, and self-correct when something breaks. On my CRUD app, it caught a schema mismatch I hadn’t spotted and rewrote the migration on its own.

The customization layer is worth turning on early. Agent Customization, which launched in late 2025, adds another layer. You can write Custom Instructions to set a persistent coding style, preferred stack, or tone for generated copy, and Skills let you save reusable build patterns (like “always scaffold a Stripe checkout”) that Agent applies automatically across sessions. If you build the same shape of thing over and over (waitlists, admin tools, internal dashboards) a good Skill turns a 40-minute build into a 5-minute one.

Where it lets you down

The pricing is the sharp edge, and pretending otherwise is malpractice. Replit’s effort-based pricing model (introduced June 2025) charges based on computational work, not per message: simple changes cost under $0.25, a full-feature build costs $4–$20+ per prompt, and Agent in Max Autonomy mode on a complex multi-hour task has run $20–$50+ per session. That’s the shape of it on a good day. On a bad day: when Agent 3 launched in September 2025, users reported dramatic cost increases, with one telling the Replit forum they typically spent $100–$250/month and blew through $70 in a night at Agent 3 launch. If you live in Auto mode and watch the meter, Core is fine. If you let Max Autonomy grind on a vague prompt overnight, you will wake up angry.

Then there’s the reliability tax. In rare cases, Agent might delete working code while trying to implement changes. Always use checkpoints before major modifications, and Replit added extra safeguards after a high-profile incident, but backups are still essential. “High-profile incident” is doing a lot of work in that sentence: the platform that hit $100M ARR in nine months also deleted a founder’s production database and billed users $1,000 in a week, and that tension defines Replit in 2026. Checkpoints work. Use them.

Agent 3 also falls apart in exactly the places you’d expect a probabilistic system to fall apart: abstract specs, ambiguous requirements, product judgment. On the trust ratings section of a test build, Agent conflated self-reported and verified ratings into a single five-star display that meant nothing. That’s the failure mode you have to watch for: the code runs, the tests pass, and the feature is subtly wrong in a way only a human reading the spec will catch. Treat Agent 3 like a fast, tireless junior developer, not a senior who can ship unsupervised.

The petty stuff grates too. In January 2026, Replit removed the free “Assistant” feature that let users ask basic code questions without using credits, and users argued that basic code comprehension, questions like “where is this behavior defined?”, shouldn’t require paid Agent queries, forcing users to either pay for simple questions or copy their code to external tools like ChatGPT for free answers. It’s a small thing that makes the whole product feel greedier than it needs to.

Should you pay for it?

Yes, if you match one of three profiles. If you’re a non-technical founder or PM who wants a working prototype live today, this is the shortest path there is. If you’re a working developer who occasionally needs to ship a small internal tool without touching your main stack, Core at $20/month pays for itself the first time you’d otherwise have spent a Saturday on Vercel + Supabase glue. If you’re a designer who wants to see a real interactive version of your idea, Agent 3 will get you there faster than any competitor I’ve tested.

Skip it if you’re a professional developer working on a serious existing codebase. Cursor lives in your repo and edits it, Replit builds a new one for you, and those are different jobs. Skip it if your organization requires strict cost predictability or governance over where code runs. And skip it if you can’t bring yourself to watch the credit meter, because you will bleed money.

The upgrade math is simple. Replit has four tiers: Starter (free), Core ($20/month annual), Pro ($95/month annual), and Enterprise (custom), with credits as the single currency covering agent sessions, deployments, and compute from one shared pool. Core includes $25/month with no rollover, Pro includes $100/month with a one-month rollover. Start on Starter to see if the vibe fits. Move to Core the moment you ship your first real thing. Only jump to Pro if you’re consistently exhausting Core’s credit pool or you need Pro’s higher-power models and 28-day database restore for a project you can’t afford to lose.

The bottom line

Replit Agent 3 is the best pure prompt-to-deployed-app tool I’ve used, and it’s not particularly close on that specific job. In an hour it does what used to take a weekend, and the platform’s growth numbers, 50M users and $525M in annualized revenue by April 2026, up 1,775% year-over-year, with a $400M raise at a $9B valuation in March 2026, reflect a product that’s genuinely working for a very large number of people. But it earns an 82, not a 91, because the credit model still bites unpredictably and Agent 3 still needs a human editor whenever the spec gets abstract. It’s a strong buy for the right buyer, a mistake for the wrong one, and either way you should keep one hand on the checkpoint button. The one to beat for prompt-to-app; not yet the one to beat for real engineering.