Replit Agent 3 Pricing: What You're Actually Paying For

We ran Agent 3 for a week and watched our credit balance evaporate in real time. The Register reported one user burning through $1,000 in a single week after the agent launched on September 10, 2025. Another spent $70 in a single night. Replit calls this “effort-based pricing.” Users on their own forum called it a shock.

The headline numbers look clean: Core is $25/month, Pro is $100/month. Both bundles include credits. But the advertised price and the real price are not the same number.

What Replit Says

Replit’s pricing page offers three tiers. Core gives you $25 in monthly credits for $25/month (or $20/month billed annually). Pro gives you $100 in monthly credits for $100/month (or $95 billed annually). Enterprise is custom.

Those credits sound like they should cover normal use. Here’s the catch: credits reflect “effort-based pricing”, which Replit launched in June 2025. The model prices tasks not by checkpoint count but by “true scope of the Agent’s work”—their phrase. Complex tasks cost more. Simple tasks cost less. On paper, this sounds fair.

In practice, users have no way to predict which bucket their request lands in before running it.

How Effort-Based Pricing Works in Practice

Before June 2025, Replit charged a flat $0.25 per checkpoint. You ran a task, it created a checkpoint, you paid $0.25. You knew the cost upfront. Repetitive, but predictable.

Effort-based pricing removed that predictability. The system now evaluates task complexity on the fly and charges accordingly. When Agent 3 launched in September, the autonomy dial cranked up—the agent now automatically spins up sub-agents for code review, security checks, and refactoring, even for small edits.

A forum user reported $32.60 for a single implementation task. Another spent $400 over two days on routine feature work. A third reported $20 for one UI redesign prompt. The variance is enormous.

The Real Bills

The Register obtained receipts from users after September 10. One paid $1,000 in a week versus a prior monthly bill of $180–200. Another paid $70 in one night when their usual monthly spend was $100–250. One user saw a 20x cost increase.

Here’s the per-task breakdown from forum complaints: $2–3 for small features, $0.50–$1.00 for simpler prompts, $30 for a single app that didn’t work. None of these included a refund for failed runs. Once the task executes, the cost sticks, regardless of whether the output was usable.

The spike hit because Agent 3 doesn’t just take your instruction and code. It spawns multiple processes—one agent plans the change, another codes it, a third reviews security, a fourth applies the fix. All four agents bill separately, each consuming effort credits.

Why Editing Existing Code Costs More Than Building New

This is where the pricing model breaks down most visibly. Building a new app from scratch triggers fewer sub-agents. The agent makes assumptions, builds toward a goal, ships it.

Editing existing code triggers a multiplier effect. The agent reads the existing codebase, evaluates potential breaking changes, plans a refactor, reviews the code for security flaws, and applies the fix. Each step is a separate agent run. Users reported $2–4 per action on existing code versus $4–6 for hour-plus builds on fresh projects.

The sub-agent multiplication makes iterative development—fixing bugs, adding features, refining UX—prohibitively expensive. Replit’s own framing calls this a feature (“reflect[s] the true scope of the Agent’s work”). Users call it a tax on iteration.

The Predictability Problem

Before you run a task, you don’t know if it will cost $1 or $100. Replit provides no pre-execution estimate. You type a prompt, hit submit, and the bill arrives after the work completes. If the task fails—the code doesn’t compile, the feature doesn’t work, the agent hallucinates—you still pay the full cost. There is no refund mechanism.

Effort-based pricing only works if effort is transparent. When it’s opaque until after billing, it’s just variable pricing with plausible deniability.

Compare this to token-based systems like GitHub Copilot, which let you see token consumption in real time. Or flat-rate tools like Cursor, which charges $20/month with unlimited use, which is predictable.

How It Compares

Cursor’s flat $20/month model doesn’t use effort-based pricing. You pay once; you use the tool. No surprises. GitHub Copilot bills by token count per request, but the token price is published upfront. Windsurf’s quota model caps requests per month and lets you know the ceiling before you hit it.

Replit’s effort-based model sits in a gap: variable pricing, opaque rules, no spend cap until your credits are gone.

Our Verdict

Replit Agent 3 is not $25/month or $100/month. It’s $25/month plus a variable overage that you can’t predict and can’t refuse after execution. For light use—a prompt or two per week—you might stay inside the bundle. For active development, especially on existing codebases, the advertised price is fiction.

If you’re planning to use Agent 3 heavily, budget 5–10x the advertised plan cost, or switch to a flat-rate tool that doesn’t penalize you for iteration.