Replit Pricing in 2026: What the Agent Actually Costs

Replit pricing at a glance

Replit pricing has two parts that people routinely confuse: a monthly subscription, and usage-based credit consumption on top of it. The subscription buys you a credit allowance and a feature tier. The Agent then spends those credits at a rate it decides per request, and anything past the allowance is billed as overage.

Here is what Replit's pricing page shows, checked on 18 September 2026:

Two details that matter and are easy to miss. First, Pro is sold in credit tiers, not as a single price: the annual view of the pricing page lists tiers at $90, $215, $425, $825 and $2,000 per month. You are picking a credit budget, not just a seat. Second, Pro replaced the old Teams plan and has no per-seat fee. Replit's own billing docs say Pro gives you up to 15 builders with pooled credits, which is a real change from the $35-per-user Teams plan that Replit sold in 2025.

The Starter plan is not on the main pricing page at all any more, but it is still documented. Replit's Starter plan doc describes daily Agent credits up to a monthly cap, one published app that goes offline after 30 days, and a "Made with Replit" badge you cannot remove without upgrading. Full build, Plan Mode, third-party connectors and additional published apps all require Core.

One conflict worth flagging: several third-party pricing trackers list Core at $25 per month on month-to-month billing rather than $20. Replit's live pricing page is the source we trust, and it shows $20 with $18 per month billed annually. If you see $25 quoted somewhere, treat it as a stale or a differently-toggled reading of the same page.

How effort-based pricing actually works

This is the part buyers search for, and it is the part that decides the bill.

Until mid-2025, Replit Agent charged a flat $0.25 per checkpoint. A checkpoint is a snapshot the Agent creates when it finishes a unit of work, and a big task would get chopped into several $0.25 checkpoints. On 18 June 2025, Replit replaced that with effort-based pricing: one checkpoint per request, priced by the time and compute the Agent consumed.

The numbers Replit published in its effort-based pricing recap on 12 July 2025 are the most useful anchor anyone has:

• Simple requests can cost as little as $0.06.
• More involved requests "will usually cost more than $0.25".
• Larger work can run to "multiple dollars depending on the size of the work".

Replit was also candid about the direction of travel. In the same post it wrote that effort-based pricing "can end up being more expensive over the lifetime of a project", and explained why: the larger your project and the longer your chat thread, the more context the Agent has to process, so the more each incremental message costs. That is a compounding cost curve, not a flat rate.

Replit's AI billing doc adds two more things that hit the bill. Plan Mode can perform billable reasoning even when it changes no code. And Agent features that call third-party APIs behind the scenes, like Claude, ChatGPT or Gemini image generation, are billed at the provider's public API rate and deducted from your Replit credits.

How Replit pricing got to where it is

Buyers keep finding contradictory numbers for Replit because the pricing has moved three times in under two years, and most of the internet is quoting whichever version it saw first. The timeline is short and it explains every conflicting figure you will run into.

September 2024 to 2025: the product reorganises around Agent. Replit's own 2025 year in review describes the platform becoming Agent-first over the course of the year, with Agent v2 in February and Design Mode in November. That matters for pricing because once the Agent is the main way people build, metered Agent consumption becomes the main way people pay.

18 June 2025: effort-based pricing replaces the flat $0.25 checkpoint. This is the change covered above, and it is the single most important one for anyone modelling a bill. Before it, the number of checkpoints told you the price. After it, nothing you can see before the request tells you the price.

September 2025: Agent 3 ships. The same year-in-review post describes Agent 3 as self-testing in a real browser and capable of "up to 200 minutes of autonomous work". Read that next to effort-based billing and the cost complaints from that month stop being mysterious. A system that can work autonomously for over three hours on one request, under a model that prices by time and compute consumed, is a system whose worst-case request price went up sharply.

24 February 2026: Pro launches and Core gets cheaper. Replit's Pro announcement, published 24 February 2026 and last updated 15 September 2026, is the post that resolves the $20-versus-$25 confusion. It says plainly that Core dropped "from $25/month" to $20 per month, that existing Core subscribers moved to the new price at their next renewal after 25 February 2026, and that the new Pro plan starts at $100 per month with tiered credits and credit rollover. It also confirms the Teams sunset: Teams subscribers were auto-upgraded to Pro at no extra cost for the remainder of their term, rolling out from 3 March 2026.

So if a comparison site tells you Replit starts at $25, it is quoting a price that stopped being current in February 2026. That is worth knowing before you use a third-party table to build a budget.

The Agent modes have already been renamed once. The February 2026 announcement introduced Economy Mode, Power Mode and Turbo Mode. Replit's current Agent modes doc lists Free Mode, Power Mode and Max Mode instead, and states directly that "Economy Mode is no longer available. Power Mode replaces it at the same price." Where an announcement and a live doc disagree, the live doc is the one to trust. The practical lesson for a buyer is that the cost controls on this product are still being reshaped, so any internal runbook you write about which mode to use has a short shelf life.

Why this matters right now: Replit has changed its metering model, its plan names, its prices and its cost-control modes inside eighteen months, and all four changes moved money. If you are about to commit to an annual term, you are betting on a pricing structure that has not yet held still for a full year.

What a realistic monthly bill looks like

Replit does not publish a typical bill, so the model below is built from Replit's own per-checkpoint figures above. These are illustrations with the assumptions stated, not measured averages.

The blended-cost assumption is the whole ballgame. At $0.35 per request, a solo developer on Core lands near the price of a normal SaaS seat. At $0.90, the same person pays more in overage than the subscription costs. Nothing about the plan changed between those two rows. What changed is the size of the codebase and the length of the chat thread.

Note also that credits cover more than Agent. Per Replit's usage-based billing doc, the same credits pay for publishing outbound data transfer, Autoscale compute units, request counts and production database usage. A busy published app eats the same allowance your Agent does.

If you run out mid-month, Replit sells credit packs: $100 for $100, $300 for $290, $500 for $480 and $1,000 for $950, documented in Managing Your Spend. They expire six months after purchase and do not roll over.

Core or Pro: the break-even point, worked out

The most common real question is not "what does Replit cost" but "am I on the wrong plan". You can answer that with arithmetic on the published prices, and the answer is a single number.

Assumptions, stated so you can change them: annual billing on both plans, Core at $18 per month with $20 of included credits, Pro at $90 per month with $100 of included credits, and overage billed at face value once the included credits are gone. This is arithmetic on Replit's published pricing, not a measurement of anyone's account.

• On Core, once your usage passes the $20 allowance, your monthly total is $18 plus everything above $20. At $60 of usage you pay $58. At $92 of usage you pay $90.
• On Pro, anything up to $100 of usage costs a flat $90.

The two lines cross at about $92 of monthly usage. Below that, Core is cheaper. Above it, Pro is cheaper and gets relatively cheaper the further you go, because Pro's credits are included rather than billed as overage.

Three things shift that break-even in Pro's favour before you reach $92:

1. Credit rollover. Core credits reset at the start of each billing cycle and do not carry over. Replit's pricing and plans help doc says Pro includes rollover, with unused credits carrying over for two months. If your usage is spiky, a plan that banks a quiet month is worth more than the headline price difference suggests. Note a source conflict here: the February 2026 announcement said unused Pro credits roll over for one month, while the current help doc says two. The live doc is newer, so that is the one to plan against, and it is worth confirming in your own billing settings before you rely on it.
2. Parallel Agent tasks. The same doc gives Core 1 parallel Agent task and Pro up to 10. If two people are waiting on one queue, the cost of the cheaper plan is not only money.
3. Headcount. Pro covers up to 15 builders with pooled credits and no per-seat fee. At three or more people, Pro is usually cheaper than three Core subscriptions before you consider usage at all.

The honest counterweight: Pro is sold in credit tiers, so "moving to Pro" means choosing a credit budget of $100, $215, $425, $825 or $2,000 per month. Picking a tier you do not fill is the most expensive mistake available on this pricing page, and rollover only softens it for two months.

Where the cost surprises come from

The complaints are dated, public and specific, which makes them useful rather than anecdotal.

The Register reported on 18 September 2025 that users were hit with sharp cost increases after Agent 3 launched on 10 September. One user told the publication that in a single week the Agent charged roughly $1,000, against a prior norm of $180 to $200 a month for similar work. The same user's read on why: editing an existing codebase is far more expensive than building something new, because the Agent spawns subagents to review, plan, security-check and fix across thousands of existing lines. Another user quoted in the piece said a single prompt cost $20 and rewrote the app's UI unasked.

InfoWorld covered the same episode on 19 September 2025, citing an Agent 3 feedback megathread on Reddit and an analyst's explanation that Agent 3 spawns subagents to refactor code even on minor edits, so "the refactoring is more expensive than original creation".

Replit has acknowledged rollout problems directly. In the July 2025 recap it disclosed a 11 July pricing incident where checkpoint charges were miscomputed for about 6% of paying users, refunded those charges, and issued $10 in credits to everyone on the new model. Whatever else you conclude, the company is not pretending the transition was clean.

The pattern across all of it is the same: cost scales with codebase size and context length, not with how big the change felt to you. A one-line fix in a large old project can cost more than building a new app from scratch.

Worth separating two kinds of evidence here, because they carry different weight. The $1,000-week and the $20 prompt are single users' accounts of their own invoices, reported by a publication but not independently auditable. The mechanism underneath them is structural and documented: Replit itself states that longer threads and larger projects cost more per message, and it states that one request can now trigger up to 200 minutes of autonomous work. You should treat the specific dollar figures as illustrative and the mechanism as fact.

Five billing surprises that are structural, not bad luck

These are not edge cases. Each one follows from how the product is documented to work, and each one has caught people who read the pricing page carefully.

1. Free Mode is not an unlimited mode. Replit's pricing and plans help doc says Free Mode provides no-cost Agent work within each plan's allowance, and that "the Core and Pro allowances reset every five hours and have weekly limits". Two ceilings, not one. A builder who plans a long Saturday session on Free Mode can hit the weekly limit and find that the next request is a paid one.

2. Your app's success competes with your Agent budget. Credits are one pool. Replit's usage-based billing doc gives Core a 100 GiB monthly allowance for publishing outbound data transfer, and bills Autoscale deployments in Compute Units, where one RAM second counts as 2 units and one CPU second counts as 18. If a published app gets traffic, it draws down the same balance your Agent is spending. The month your project works is the month your credits disappear fastest.

3. Plan Mode can bill even when nothing changes. Replit's AI billing doc says Plan Mode can perform billable reasoning in the paid path even when it does not change code. Planning is still usually the cheaper path, but "I was only planning" is not a defence against a charge.

4. Credit packs expire, and auto-reload buys more without asking again. Per Managing Your Spend, credit packs expire six months after purchase, do not renew, and do not roll over once expired. Replit applies the pack that expires earliest first, which is the right default. Auto-reload, if you turn it on, buys a new pack automatically when your balance runs low, which is convenient during a build and expensive if you forget it is on.

5. Third-party model calls bill at the provider's rate. Image generation and similar features that call Claude, ChatGPT or Gemini behind the scenes are billed at those providers' public API rates and deducted from your Replit credits. That cost is real, it is not a Replit markup, and it does not show up in any "price per checkpoint" estimate you build.

How to keep a Replit bill predictable

Replit ships the controls; they are mostly off or unused by default.

• Set a budget limit, not just an alert. Settings, then Account, then Billing. A hard cap is the only thing that actually stops runaway spend.
• Use Plan Mode before Build Mode. Replit's own guidance is that planning conversations burn fewer tokens than code generation, and you get to approve scope before the expensive part starts. Remember that Plan Mode can still bill on the paid path.
• Start new chats on long threads. Replit says explicitly that longer message threads make each incremental message more expensive. This is the single cheapest habit change available.
• Pick the mode deliberately. Free Mode covers everyday work inside your allowance. Power and Max cost more per request and should be reserved for work that genuinely needs them.
• Check the usage page weekly. Usage data can take up to 30 minutes to appear, so a daily glance during a heavy build is not paranoia.

Four more levers that are documented but rarely used:

• Default to Power Mode, not Max. Replit's Agent modes doc describes Power Mode as "optimized for cost" and recommends it for maximising the number of prompts per credit, with Max Mode reserved for complex, production-grade work. Most teams set the most capable mode once and never change it back, which is the most expensive default available.
• Cap spending to only what you have pre-bought. This is the sharpest trick in Replit's own docs and almost nobody uses it. Usage limits and budgets apply after any credit packs are consumed, so if you set your limits to $0.01, you are effectively restricted to the credits already sitting in your pack. That converts a metered bill into a prepaid one, which is the format finance teams can actually approve.
• Give Routines a per-run budget on purpose. The Agent modes doc notes that Routines do not run in Free Mode. They run in Power or Max and require a per-run budget. Anything you schedule is recurring spend, so set that number deliberately rather than accepting a default on something that will run every day.
• Filter the usage page rather than staring at the total. The usage page can be filtered by date, resource, project, workspace, group or member. On a team plan with pooled credits, the total tells you that you have a problem and the filters tell you where it is.

One habit worth more than all of the above: decide in advance what a request is allowed to cost, and scope it to that. Effort-based billing prices the work the Agent chooses to do. The only reliable way to control the price is to control the size of the ask.

Who Replit is wrong for

Replit is genuinely good at a specific job, and the honest version of a pricing article says who should walk away. These are all direct consequences of facts above, not opinions about quality.

Teams that need a committed number for procurement. If someone has to sign off a fixed annual figure, effort-based billing is a bad fit by design. The price is set after the work by a system that decides how much work to do. You can cap it with budgets, but a cap is a ceiling, not a forecast, and a cap that stops your builders mid-project has its own cost.

Anyone maintaining a large legacy codebase. This is the worst case for effort-based pricing and it is the exact case users complained about in September 2025. Cost scales with how much context the Agent processes, and a mature codebase is all context. If most of your work is small edits to a big existing system, the pricing model is working against your usage pattern on every request.

People who want the AI inside their existing editor and toolchain. Replit is the environment, not a plugin. You get the hosting, the database, the deployment and the collaboration along with the Agent, and you pay for all of it through one credit pool. If your code lives in a repository your team already runs, you are paying for infrastructure you will not use.

Anyone counting on Starter as a lasting no-cost option. The Starter doc is clear that you get one published app that goes offline after 30 days, a "Made with Replit" badge, Lite build only, and no third-party connectors. It is a way to try the product properly. It is not a plan you can run something real on. Note also that Replit does not offer a time-limited trial of the paid plans, so Starter is the evaluation path.

Cost-sensitive teams with steady, high-volume usage. Metered billing rewards spiky usage and punishes constant usage. If five people use the Agent all day every day, model the annual number at your real request volume before you commit, because a per-seat competitor may simply be cheaper at that shape.

Who it is right for is equally specific: solo builders and small teams who want building, hosting and deployment in one place, non-engineers shipping internal tools, and anyone prototyping new projects rather than maintaining old ones. For those users, the same pricing model that punishes large-codebase editing works in their favour, because a small project is cheap context.

Replit pricing vs Lovable, Bolt.new and v0

The nearest alternatives all price the same job differently, and the units do not convert. Everything below was read from each vendor's own live pricing page on 18 September 2026.

Four things fall out of that table.

Only v0 denominates its unit in money. A v0 credit is a dollar of credit, so "how much is left" and "what will this cost" are the same question. Replit's checkpoints, Lovable's credits and Bolt's tokens all require a conversion you cannot perform in advance.

Lovable is the only one that publishes what individual actions cost. Its pricing page lists worked examples: "Make the button gray" at 0.50 credits, "Remove the footer" at 0.90, a prompt adding authentication with signup and login pages at 1.20, and "Build me a landing page, use images" at 1.70, with Plan Mode at a flat 1 credit per message. Lovable's credits doc adds the top-up rates, at $15 per 50 credits on Pro and $30 per 50 credits on Business, which means a credit costs twice as much on the more expensive plan. That is unusual and worth reading carefully before upgrading.

Bolt is the only one whose entry paid plan rolls unused capacity forward by default, per its pricing page. Replit reserves rollover for Pro. Lovable offers it on Pro. Rollover matters more than most buyers expect, because every one of these products produces uneven months.

Replit is the only one with no per-seat fee at the team tier. Pro covers up to 15 builders on pooled credits. Bolt Teams is $30 per member, v0 Business is $100 per user, and Lovable Business is a flat $50 for the workspace. For a five-person team, that difference dominates the comparison before anyone runs a single request.

The comparison people actually want, which no table can give them, is the total for a month of their own work. None of these vendors publishes it, because none of them can. What you can do is pick the product whose failure mode you can live with: a hard token ceiling that stops you, or a metered bill that does not.

How Replit pricing compares to other AI coding tools

Replit is not an outlier in direction, only in mechanism. The whole category moved from flat seats toward metered consumption during 2025 and 2026, which is why Windsurf's pricing is worth reading alongside this one, and why we keep a running comparison of enterprise AI pricing models. Claude Code's pricing is the closest thing to a like-for-like read on how a terminal-first coding tool prices the same work, and our roundup of the best AI for coding covers where each tool is actually strong.

What makes Replit distinct is that the unit of billing is opaque at purchase time. A per-seat tool tells you the price before you use it. A per-token tool at least lets you estimate. Effort-based checkpoints price after the work is done, by a system that decides how much work to do. That is a genuine strength when the Agent is efficient, because you pay $0.06 for a trivial fix instead of $0.25. It is the whole problem when the Agent decides a one-line change warrants a full-codebase review.

Worth saying plainly: Replit's model is defensible. The company's argument that user cost should track its own compute cost is honest, and it published the reasoning rather than hiding it. The gap is forecastability, not fairness. If you want the wider view of what a seat of AI tooling should cost across a company, how much enterprise AI should cost per user puts these numbers in context.

Where an MCP layer fits alongside Replit

If your actual question is "how do I stop my AI coding tool from re-reading the whole codebase to figure out context", that is a different problem from which Replit plan to buy. Replit builds and hosts the app. A Model Context Protocol layer is what gives AI coding tools structured access to the systems around the code: tickets, docs, CRM records, internal wikis.

Coworker MCP connects AI tools like Claude Code, Cursor and Windsurf to 50+ apps through one governed connection, at published pricing: Pro at $29.99, Max at $149.99, and a custom Enterprise tier. It does not replace Replit and does not build apps. It gives the tools you already use a single, permissioned way to reach company context, so less of a request's budget goes into rediscovering things your systems already know. If you are evaluating that layer, how to choose MCP servers covers what to check before you commit, and MCP versus a direct API integration explains when the protocol is worth the extra hop.

Book a demo to see how Coworker MCP fits alongside the coding tools your team already pays for.

Frequently asked questions

How much does Replit cost per month?

Replit's pricing page lists Core at $20 per month, or $18 per month billed annually, and Pro at $100 per month, or $90 per month billed annually. A no-cost Starter plan exists and is documented separately, and Enterprise is custom-quoted. Those subscription prices include a credit allowance; Agent usage beyond it is billed on top.

What is effort-based pricing on Replit?

It is the model Replit introduced on 18 June 2025 to replace flat $0.25 checkpoints. The Agent now creates at most one checkpoint per request and prices it by the time and compute the request consumed. Replit's own figures put simple requests as low as $0.06 and larger requests at multiple dollars.

Why did my Replit bill go up without me changing plans?

Almost always because the project got bigger or the chat thread got longer. Replit states that the larger the project and the longer the message thread, the more each incremental message costs. Editing a large existing codebase is the most expensive pattern reported, because the Agent reviews far more code than the change itself would suggest.

Can I cap what Replit charges me?

Yes. Replit supports usage alerts and hard budget limits under Settings, then Account, then Billing, plus credit packs for planned spikes. The caps are not applied by default, so they have to be set deliberately before a heavy build rather than after a surprising invoice.

Does Replit still have a Teams plan?

No. Replit's billing docs state that the Pro plan replaced the former Teams plan, which was sold at $35 per user per month. Pro has no per-seat fee and covers up to 15 builders with pooled credits and shared budget controls.

Is Replit cheaper than other AI coding tools?

At the subscription level the base prices are comparable to peers. The difference is variance: metered agent billing means two teams on the same plan can pay very different amounts in the same month. Compare on expected total spend under your own workload, not on the sticker price, and read the overage terms before committing to an annual plan.

When should I move from Replit Core to Pro?

On published annual prices, Core costs $18 plus anything above its $20 credit allowance, while Pro is a flat $90 up to its $100 allowance. Those cross at roughly $92 of monthly usage, so above that level Pro is cheaper. Move sooner if you have three or more builders, since Pro covers up to 15 with no per-seat fee, or if your usage is uneven, since Pro credits roll over and Core credits do not.

Do unused Replit credits roll over?

Not on Core. Replit's help documentation says Core credits reset at the start of each billing cycle, while Pro includes rollover with unused credits carrying over for two months. Separately-purchased credit packs expire six months after purchase, do not auto-renew, and are consumed earliest-expiring first.

Why is Replit charging me when the app is just running?

Because credits pay for hosting as well as for Agent work. Replit bills publishing on outbound data transfer, Autoscale compute units and request counts, plus production database usage, out of the same credit balance. Core includes a 100 GiB monthly outbound data transfer allowance. A published app with real traffic will draw down the balance you were saving for building.

Is Replit's Starter plan enough to evaluate the product?

For a first look, yes, and it is the only evaluation route since Replit does not offer a time-limited trial of the paid plans. But the limits are real: one published app that goes offline after 30 days, a "Made with Replit" badge, Lite build only, and no third-party connectors. If your evaluation depends on connecting to Google, Stripe or any other external service, you need Core before you can judge the product properly.

Related reading

• Windsurf Pricing Explained
• Claude Code Pricing
• Best AI for Coding
• Enterprise AI Pricing Compared
• How Much Should Enterprise AI Cost Per User?
• How to Choose MCP Servers for Your Team
• MCP vs API
• What Is MCP?