Replit Agent 3 in 2026: What Replaced It, Limits and Costs

If you came looking for a Replit Agent 3 review, start here: Agent 3 is no longer the version you would be using, because Replit replaced it with Agent 4 on 11 March 2026. The parts people liked carried over, namely an agent that tests its own work in a real browser and bills by effort. Almost everything around them changed, including the modes, the pricing plans and how unattended work runs.

This is not a hands-on review. We have not run long autonomous sessions to failure, so there are no burn rates or size ceilings of our own here. It is built from Replit's launch posts, docs, pricing page and changelog, checked October 2026: the Agent 3 launch post, the Agent 4 launch post, what changed from Agent 3 to Agent 4, pricing, and the docs on Agent modes, AI billing, the task system, App Testing, managing spend and publishing costs.

Replit sits on the same shelf as Lovable and Bolt: describe an app, get a running one. Where it differs is that the editor, database, login and hosting are one product. If you are still deciding whether a builder belongs in your stack at all, our guide to building an AI SaaS in 2026 covers that decision first.

Replit Agent 3: what it promised, and where each promise stands

Agent 3 launched on 10 September 2025 with autonomy as the headline. Replit called it 10x more autonomous than the version before. Here is each launch claim next to what the docs say now.

Is a Replit Agent 3 review still worth reading?

Only for history. A review written between September 2025 and March 2026 describes Design Mode, fork-and-merge collaboration, Economy Mode and an Autonomy Level setting. None of those exist under those names now. Before you trust any verdict, check its date against the timeline above.

Replit's own comparison post says old projects still work and nothing was deleted. It also says the Design Canvas, shared collaboration and parallel task execution are available in new projects, so an older project gets less of Agent 4 than a fresh one.

Autonomy today: how does an unattended build run?

The 200-minute headline has given way to a task system. A task is a unit of work Agent can plan and execute, and background tasks run in isolated copies of your project, so the main version stays untouched until you apply the result.

Three mechanisms do the work that "autonomy" used to describe:

• Background tasks. Tasks move across a board with four columns: Drafts, Active, Ready and Done. When two tasks touch the same files, Agent resolves the conflict as it applies them.
• App Testing. The docs say it gives Agent extended autonomy to work for longer without human intervention. Agent decides when enough has changed to test, and you can replay the session as video afterwards.
• Checkpoints. Open History in the Agent panel and each checkpoint has Rollback here and Changes controls. The rollback dialog lists the files, Agent memory and tasks affected, with a Database checkbox.

That last checkbox deserves a second look before you click. A rollback that includes the database is the right call after a bad migration and the wrong one if real users have written data since.

Which Replit Agent mode should you use?

Start in Free Mode and move up only when a task needs it. That is Replit's own advice, and the modes are built around it. You pick one from the Agent settings dropdown in the prompt box.

Two details matter for cost control. Replit says it lets you know before a request moves from Free Mode to a mode that incurs usage costs, so a paid build should never start by accident. And these settings belong to each user, not to the project, so on a plan with pooled credits, a teammate working in Max Mode draws on the same pool at a different rate from you.

Limits: what do the docs say it will not do?

Replit documents its limits more plainly than its launch posts suggest. These are the ones that decide whether the Agent fits a project.

Two of these catch people after they have committed. App Testing covers two kinds of web app, so a mobile build or another framework gets no self-testing. And the free plan's published app goes down after 30 days until you publish it again, which is fine for a demo and not for anything you have told customers about.

Replit Agent pricing (checked October 2026)

Plan prices are the smaller half of the bill. Agent uses effort-based pricing: each completed request produces one checkpoint, and a simple fix costs less than a full feature. Replit asks for confirmation before a paid action starts, and you can hover over the usage icon in the Agent tab to see what each checkpoint cost. It publishes no price list per task; the AI billing page points to an interactive demo of sample app costs.

Credits also pay for running the app, not only for building it. Publishing costs come out of the same monthly credits, then bill by usage:

The controls are real. Under Settings, Account, Usage you can set a budget limit, and Replit says usage-based services are suspended until the budget is raised or the next billing cycle begins. Credit packs run from $100 to $1,000 with small discounts on the larger ones, and expire six months after purchase. The docs' own advice for spending less is short: plan before building, start in Free Mode, and approve only the changes you want.

Is Replit Agent worth it, and who does it serve?

It is worth it for people who value one product over a stack, and who can live with a bill that moves. Replit gives you the builder, the database, login, storage and hosting without leaving the tab. The price of that convenience is usage-based billing on both the building and the running.

If a fixed allowance matters more to you, our Lovable vs Bolt.new comparison covers two builders that sell credits or tokens by the month. If you would rather own the stack than rent all of it, our AI SaaS Builder program covers that route: Supabase, Next.js and Vercel, with lessons on Cursor and Claude Code.

How to evaluate Replit Agent yourself in a weekend

No review, including a hands-on one, can tell you what your project will cost. A weekend with a budget limit can. Each step below uses a control the docs describe.

Before you publish: Replit's own security checklist

An agent that builds and hosts the app does not remove the launch checks. Replit publishes a security checklist in three parts, and it reads like a list of things to ask the Agent for by name:

• Front end: validate and sanitise input, keep sensitive data out of the browser, and never expose API keys in frontend code. Replit Secrets is the place for keys.
• Back end: require authentication on every API endpoint, check permissions before any action runs, and use parameterised queries or an ORM. The checklist notes that Agent uses ORMs by default and recommends Replit Auth.
• Habits: keep dependencies updated, keep internal details out of error messages, and rate limit every endpoint, authentication ones most of all.

Those are the same launch checks as for any AI-built app. Our guide to taking a vibe-coded app to production walks through the full pass, and for the terminal-first alternative to a hosted builder, see the Claude Code guide.

Replit Agent 3 and Agent 4: FAQ