Charlie Meyer's Blog

Replit's Existential Problem

Disclaimers

Background

Replit has always been a site for writing code online, and eventually transitioned to a platform for hosting code, then a fully AI powered “vibe coding” tool. I argue that this transition has put Replit in a very precarious position.

Replit started as a great idea

On the first available entry of replit.com on the Wayback Machine, Replit described their site as such:

“The repl.it project is an attempt to create an online environment for interactively exploring programming languages. It provides a fully-featured terminal emulator and code editor, powered by interpreter engines for more than 15 languages.”

This was posted in 2015, and at the time Replit was really unique for providing these features.

For the years following launch, Replit focused on serving novice programmers learning to code, making it easy for them to get started and for them to share their work. In 2019, their landing page reads:

Stop wasting time setting up a development environment.

Repl.it gives you an instant IDE to learn, build, collaborate, and host all in one place

The key word here is learn. Replit had specific offerings for education, and marketed to teachers as a solution for running their coding classes in the cloud. As of 2019, they had a paid plan for schools. For about $10/student/year, teachers had access to pro features for running their classrooms.

I personally started teaching computer science in 2021, and Replit was an invaluable tool. I used it every single day in my classes. It was how students wrote 100% of their code. The convenience and power of Replit was incredible.

Transition to AI

By 2023, after the launch of ChatGPT, Replit began to transition to AI powered programming. They dropped paid classroom plans and began selling subscriptions that gave access to Ghostwriter, their AI coding assistant (similar to Github Copilot). In October 2023, their landing page reads:

Idea to software, fast.

Build software collaboratively with the power of AI, on any device, without spending a second on setup

Over the course of 2023, Replit pushed AI features to every user by default, including students still using Replit in the classroom. For teachers, this was catastrophic. Shortly after this rollout, Replit deprecated their education product, and by fall of 2024, deleted all content created by teachers in their old Replit classroom plans. (This is a really bad thing to do to people and maybe says something about how Replit thinks about customers)

By Fall 2024, Replit’s page reads:

Build software faster

Replit is an AI-powered software development & deployment platform for building, sharing, and shipping software fast.

Replit is now not about learning, but about “shipping”, “fast”.

The Problem with “Professional Replit”

Browser based IDEs like Replit are great to lower the barrier to entry to software development, for example, for a learner just getting started. But at every turn, a browser based software development platform will be slower and less powerful than alternatives built for the desktop.

If you’re serious about software development, you will use industry standard tools like VS Code. If you’re into AI, then adding Copilot to VS Code, using Cursor, or command line tools like Claude Code are the ways to go. Collaboration is handled with Github, or maybe Gitlab if you don’t like Microsoft.

So, who is a browser based coding platform for, exactly? There are endless tools for professionals that run locally and provide better performance and more features.

Enter the “Vibe Coder”

Andrej Karpathy coined the turn vibe coding in February 2025 with a post on X. On February 3rd 2025, Replit CEO Amjad Majad quote tweeted his post:

Vibe coding is already here.

75% of Replit customers never write a single line of code!!

This is not a good statistic to be sharing, and honestly can’t be true if people are building software of any seriousness on Replit. No matter how enthusiastic you are about AI, to get a real project off the ground you’ll need to write at least a couple of lines of code.

By February 2025, Replit’s marketing page fully embraced this idea:

Idea to app, fast. Create beautiful, modern web applications at the speed of thought. Describe what you need and Replit's AI Agent builds it for you.

Let’s take a snippet from Karpathy’s original tweet:

“Sometimes the LLMs can't fix a bug so I just work around it or ask for random changes until it goes away. It's not too bad for throwaway weekend projects, but still quite amusing. I'm building a project or webapp, but it's not really coding - I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.”

The key words here are “throwaway" and "mostly”. Karpathy describes a real use case for AI: rapidly prototyping ideas for fun or writing small pieces of software that have little long term value (throwaway).

Who is Replit for

So, who is Replit, in particular, for?

There are two scenarios:

To me, A makes no sense. Developers take pride in crafting their setups, and know how to deploy their software to cloud platforms like AWS. B simply doesn’t work for anything other than Karpathy’s throwaway software. And if it does, that non-technical user will certainly be asking more of the AI and causing more spend.

Inexperienced vibe coders are expensive

Based on the economics of today’s AI models, it is not possible to sustainably provide a flat fee subscription for a service like Replit. AI Agents, especially those working on large code bases need large context windows and burn a ton of tokens. Reasoning models are often used, adding to the costs. Whatever subscription price you’d charge would have to cover the cost of the average user plus some margin, and would simply be too high of a bar for entry. (This is an estimation, but Replit doesn’t do this, so I assume they agree).

Replit has a usage based pricing model for its agents. For the non-technical user, this is not a good experience. According to the very skewed demographic of Replit’s subreddit, this is actually an extremely bad experience.

The non-technical user does not know how to prompt an agent effectively, and every message has an unpredictable cost, since it’s impossible ahead of time to know how many tokens a given message might cause an agent to use.

If 75% of users of Replit truly don’t write a line of code, they’re using an unpredictable amount of money to implement any small change to their project, with costs that quickly add up.

Where we are today

Finally, let’s look at today’s landing page.

Turn your ideas into apps

What will you create? The possibilities are endless.

Replit cites some success stories on their landing page. Some companies seem to have built internal tools to save money or become more efficient using Replit. Let’s take them at face value. Replit might be a good prototyping tool.

This prototyping tool is valued at $3 billion on $150 million in annualized revenue. From Amjad in the post announcing their most recent raise:

“We were the first to make vibe-coding a reality for individuals with no prior coding experience across the world. With the raise and our new AI Agent, we are positioned to supercharge customer traction to become the standard for enterprises. The future is exciting with millions — if not billions — of people bringing their ideas to life with a few clicks.”

What we don’t know is:

These are Replit’s real issues.

In an alternate version of history, Replit could have turned into the platform for teaching and learning programming, but that’s not a billion dollar business.

To be fair, I don’t know what else Replit could have, or should have done (except not delete teacher data). They are under pressure from investors to provide returns. Depending on developments in the cost of inference and power of AI models, there is a chance that Replit’s business does work out and this might have been their best option for a pivot. I’m not convinced that their current path keeps them as a billion dollar business for long, but we’ll see.