Build · founder · 7 min read

Replit made most of your AI usage stop counting. Here's the catch.

Replit's Free Mode runs everyday Agent work on OpenAI's cheap Luna model at no credit cost. What it changes for your bill, and where the meter still runs.

Published August 19, 2026 ·

replit openai pricing credits model routing budget

Replit launched Free Mode today: a new default for Core and Pro subscribers where everyday Agent work — chatting, planning, small tasks — runs on OpenAI’s GPT-5.6 Luna and doesn’t draw down your credits. Replit says Core subscribers can now create “up to 30X more” for the same $20 a month, plus up to 30 hours of chat.

The announcement went to Fortune as an exclusive and landed alongside an OpenAI post, which tells you this is a partnership milestone as much as a feature. Replit’s president of AI, Michele Catasta, called making it the default “radical.” Both companies say more joint launches are coming.

The name is doing some work. Free Mode is not free. You still need a $20/month Core or $100/month Pro subscription. What’s free is the metering on a specific class of work.

Why this is a bigger deal than it sounds

The single most common complaint about this entire category — across Replit, Lovable, Bolt, and everyone else — isn’t output quality. It’s the meter.

You describe a feature. The agent takes a swing. It’s wrong. You explain what’s wrong. It takes another swing. Somewhere in that loop you stop thinking about the product and start thinking about what each retry costs. This site has documented that anxiety repeatedly, including Replit’s own Agent 3 credit blowups in April, where users reported $1,000 weeks against $200 monthly norms.

Free Mode attacks the psychological half of that problem directly. One of Replit’s early-access quotes puts it better than the marketing does: “My choices are bolder because I’m not worrying about burning tokens.” That’s the actual product here. Not more compute — less hesitation.

For a PM or founder, the practical unlock is the part of the work that always felt wasteful to pay for: thinking out loud. Sketching the data model. Asking “what would break if I did it this way.” Renaming things. Reading back what you already built. That work is genuinely valuable and it was previously billed at the same rate as shipping a feature.

What made it possible

OpenAI cut GPT-5.6 Luna’s API price by 80% on July 30. Catasta says that cut is what made Free Mode viable, and OpenAI’s Thibault Sottiaux told Fortune the reduction came from efficiency gains in running the model, not from new compute supply.

That distinction matters if you’re forecasting. Price cuts driven by hardware gluts reverse when the glut ends. Price cuts driven by inference efficiency tend to stick.

It’s also the clearest example yet of a pattern this site has been tracking all year: the model layer is splitting into cheap and expensive tiers, and the tools on top of it are quietly routing between them on your behalf. Replit is just the first builder to make the cheap tier the default and tell you about it in the pricing.

Where the meter still runs

Read the fine print before you rewrite your budget.

It resets every five hours

Free Mode has usage limits that reset on a five-hour cycle. Pro gets higher limits than Core. Replit hasn’t published the actual numbers, which is the least satisfying part of the announcement — “unlimited until a limit we won’t name” is a hard thing to plan against. Two early-access users said they couldn’t find the ceiling, which is encouraging and also not a number.

Real builds still cost credits

When Agent decides a task needs more than Luna can handle, it routes to GPT-5.6 Sol for the duration and then drops back to Free Mode. That routed work is Power Mode or Max Mode, and those consume credits normally.

So the honest framing is: conversation got free, construction didn’t. If your usage is 80% building and 20% planning, your bill barely moves. If it’s the reverse — and for a lot of non-technical founders, it genuinely is — this is a real cut.

Two related renames: Economy Mode is now Power Mode (same cost, same or better performance), and Max Mode is the high-reasoning tier for large changes.

You don’t pick the model anymore

Free Mode is Luna-only. You can normally choose among several models on Replit; in this mode you can’t. The agent decides when to escalate. That’s the routing trade this site keeps flagging — you gain simplicity and lose the ability to say “no, use the cheap one, I know what I’m doing.”

Hosting is a separate bill

None of this touches Replit Cloud usage. Those rates dropped sharply on August 1 and are still metered on their own. Free Mode is about making things, not keeping them live.

What to actually do

If you already pay for Replit Core or Pro: nothing. It’s on by default. But spend a week watching where Agent escalates out of Free Mode — that’s now a visible signal of which of your tasks are genuinely expensive, which is information you didn’t have before.

If you left Replit over credit burn: worth another look, with a caveat. The thing that burned you was Agent 3 spinning up autonomous subagents on complex builds. That work still escalates and still costs. Free Mode fixes the small stuff, not the thing that produced the $20 prompts.

If you’re comparing builders right now: this changes the $20-tier math meaningfully. Replit Core at $20 with unmetered everyday work compares differently against Lovable and Bolt’s credit tiers than it did last week. But don’t switch tools over one pricing change. Pricing in this category has moved four times this year already, and the cost of migrating a half-built app will outrun the savings almost every time.

If you’re on the free tier: nothing changed for you. Despite the name.

The part worth watching

The interesting question isn’t Replit’s bill. It’s whether “cheap model by default, escalate on demand” becomes the standard shape of this product category.

The economics point that way. If a small model can handle the majority of interactions at 20% of the cost, and the tool can detect the minority that can’t, then every builder still charging frontier rates for “what should I name this table” is leaving an obvious opening. Expect Lovable, Bolt, and the rest to ship some version of this within a quarter.

The risk is the same one that comes with every router: the escalation decision is now the product. If Agent over-escalates, you pay for frontier reasoning on trivial work and never see it happen. If it under-escalates, you get Luna-quality output on a task that needed better and you find out when the build breaks. Neither failure is visible from where you’re sitting.

Watch your credit burn for a couple of weeks rather than trusting the “30X” number. Thirty times more of the cheap thing is not thirty times more of the thing you actually needed.

Sources

• Fortune: Replit taps OpenAI’s low-cost Luna model for new ‘Free Mode’
• Replit: Introducing Free Mode
• OpenAI: Replit expands access to software creation with GPT-5.6 Luna