Slower, Pricier, And Everyone's Leaving? The Lovable Exodus, Audited

Every few weeks someone posts the same thread. Different subreddit, same wording. "Anyone else feel like Lovable has got slower?" "Is it just me or is it burning credits way faster than it used to?" "I've moved to Cursor, honestly should have done it months ago." Forty comments of agreement, three of them from people who have built one landing page, and one lonely voice pointing out that they shipped a paying product last Tuesday.

This site is built in Lovable. So before anyone accuses me of carrying water in either direction: I am the guy who pays the bill, watches the meter, and swears at the screen at 1am. That is the only credential that matters here.

So let's do what the threads never do. Let's separate the three accusations — slower, more expensive, everyone's leaving — and check each one against evidence rather than vibes. Two of them are substantially true. One of them is a story people tell themselves. And the reason for all three is the same, and it is not the reason the forums think it is.

Accusation one: it has got slower

Partly true, and for once the company documents its own crime scene. Lovable runs a public status page, and it is not shy. "Lovable Agent running slower than usual" — 31 March 2026, Editor degraded. "Lovable Agent slower than usual" — 1 April 2026, resolved in 38 minutes. "Agent queue gets stuck after first message" — 3 April 2026, an admitted regression where the agent simply stopped picking up queued messages; the official workaround was to pause the queue and send messages the old-fashioned way. "Delayed chat responses and real-time updates" — 25 August 2026, roughly 1.3 hours of the editor not showing you what it was doing while it did it.

That last one matters more than its severity rating suggests, because it explains a lot of the perceived slowness. When the streaming layer lags, the agent is working and you are staring at nothing. Humans do not experience "no feedback for ninety seconds" as a UI bug. They experience it as the product being slow, and then they refresh, and then they send the message again, and now they have paid twice for the same instruction. More on that later, because that is where the money actually goes.

But there is a second, less forgivable cause, and it is architectural. The agent got smarter by getting more thorough. It reads more of your codebase before it edits. It runs typechecks. It looks at your build output. It plans. Every one of those behaviours is a straight trade of wall-clock time for correctness, and every one of them is a good trade — right up until you asked it to change a button colour and it spent forty seconds reading your database schema first.

“The agent did not get slower. It got more careful, and nobody asked you whether you wanted careful on this particular button.”

And there is a third cause that has nothing to do with Lovable at all: your project got bigger. The thing you are comparing against is the memory of month one, when your app was four files and a hero image. It is now ninety files, a Supabase schema, a payments flow and eleven routes. Of course the agent takes longer. You are not using a slower tool. You are using the same tool on a fatter problem, which is precisely what happens to every codebase ever written, including the ones built by humans who bill £700 a day.

Accusation two: it has got more expensive

This one is true, and it is true in a sneakier way than people realise, and the sneakiness is mostly about accounting rather than greed.

Until recently Lovable billed you along two axes. You had a subscription that bought build credits — the messages you send in the editor. Separately, you had dollar-denominated balances for Cloud (your hosted backend) and for the AI gateway (model calls your deployed app makes on behalf of your users). Two meters, two mental models, two things to be surprised by.

As of the August 2026 change, there is one balance. Credits now pay for all of it: Build (you, prompting), plus Run (your app existing, your app's users calling models). Pro starts around $25 a month, Business around $50, both carrying 100 credits, which puts a Pro credit near $0.25 and a Business credit near $0.50. Verified against Lovable's own pricing and credits documentation in August 2026.

Why unification felt like a price rise even where it wasn't

Lovable's line is that consolidating balances does not change the underlying cost of Cloud or AI usage, and as far as I can tell that is accurate. But that is a statement about unit economics, not about human psychology, and human psychology is where the anger lives.

• →Cross-contamination of blame: your balance now falls for three unrelated reasons. When it drops, you cannot tell from the number alone whether it was your prompting, your hosting, or a hundred strangers using the AI feature in your live app. So people blame the visible thing — the editor — for spend caused by the invisible thing.
• →Loss of the free-feeling tier: when Cloud and AI sat in their own dollar buckets, small usage felt like it cost nothing. Now every backend call nibbles the same pot you need to build with, so success at the front end is punished at the back end.
• →Your users can now spend your build budget: if your deployed app calls a model, your customers are burning the same credits you need to fix bugs. That is a genuinely new failure mode and almost nobody plans for it.
• →The meter reads in credits, your brain reads in pounds: a 12-credit code review sounds cheap. Three pounds for a code review that a human would have done in the shower sounds different.

The debugging loop is the real bill

Here is the part every honest review lands on, including the ones written by people who like the product. The expensive thing is not building. The expensive thing is the loop where the agent fixes its own mistake, breaks something else, and you pay for the round trip. There are documented reports of 96 credits vanishing into one unresolved issue, and of agent mode consuming 12 credits on a single code review of a large project. Independent 2026 reviews put the same finding in politer language: metered credits drain in debugging loops where you pay to fix the AI's own errors, and there is a complexity wall past the prototype stage.

The community-consensus number — repeated across enough independent Reddit round-ups that I'll treat it as folk wisdom rather than data — is that Lovable reliably gets a project about 70% done, and the remaining 30% is where the money goes. That is not a slur. That is an extremely useful thing to know before you start, and Lovable's marketing will never tell you, because "70% of an app in an afternoon" is a worse headline than "an app in an afternoon".

“Nobody is angry about the price of building. They are angry about the price of un-breaking.”

Accusation three: everyone is leaving

This is the one that does not survive contact with evidence, and I say that as someone who would rather enjoy writing the opposite.

Lovable told TechCrunch in June 2026 that it had passed $500 million in annualised revenue, with around a million new projects created per week. That followed $400 million in February, which itself followed $300 million in January — a jump of more than 30% in a single month, confirmed by Lovable's own chief revenue officer to Business Insider in March. The company sits on a reported $13.3 billion valuation, roughly 26.6x annualised revenue — which, note, is a lower multiple than at its Series B, because revenue grew faster than the price tag. Analysts poking at that valuation in August 2026 concluded the open question is gross margin, not demand.

You do not print those numbers during an exodus. You print them during a land grab.

So what are people actually seeing when they see an exodus? Three things, none of which is collapse:

• →Graduation, not defection: the most-cited "why I left Lovable" destination is Cursor at $20/month, and the reason given is almost always the same — the prototype worked, real users arrived, and the work moved from generating features to maintaining them. Builder.io put it best in January 2026: usually you want a Lovable alternative because you succeeded at Lovable.
• →Survivorship bias in reverse: nobody posts "month nine, still shipping, bill is boring". They post the breakup. Forums are a record of friction, not of usage, which is why every mature platform's community reads like a hospice ward.
• →Genuine churn at the bottom: a million new projects a week means an enormous number of people trying it, abandoning it, and being counted by nobody. That is not an exodus from Lovable. That is the ordinary mortality rate of ideas.

What is actually happening, in one paragraph

Lovable stopped being a demo generator and started being a platform, and platform economics are less magical than demo economics. The agent became more thorough, so it became slower. The billing became unified, so success at runtime now competes with iteration at build time. And the users who got furthest hit the point every builder hits, where the constraint stops being "can I make this" and becomes "can I afford to keep changing this". None of that is a scandal. All of it is a phase transition, and phase transitions always look like decline from the inside.

How to make it cheap and fast again, since nobody in those threads will tell you

I run this site on the platform in question, so this is not theory. These are the behaviours that moved my own bill, in descending order of how much they mattered.

• →Stop paying for retries you caused. If the editor lags, do not resend the message. The agent is almost certainly still working. Re-sending is the single most expensive habit on the platform and it is entirely self-inflicted.
• →One instruction per message, but the whole instruction. Batching six unrelated changes into one prompt guarantees the agent reads half your codebase. Splitting one change into six vague follow-ups guarantees you pay six times to arrive at the same place. Precise and complete beats both.
• →Say what not to touch. "Change only the header component, leave routing and styles alone" removes an entire category of collateral damage — and collateral damage is what the 30% actually is.
• →Discuss before you build. Planning and discussion are drastically cheaper than generate-then-repair. A five-minute conversation about the data model beats three days of migrations.
• →Watch Run separately from Build. If your deployed app calls a model on the user's behalf, cap it, cache it, and put it behind a login. Otherwise strangers are spending your build budget and you will blame the editor.
• →Know your exit before you need it. Push to GitHub from day one. The code is React and Supabase — genuinely portable, genuinely readable, and the presence of a real exit is what stops "more expensive" from becoming "trapped and more expensive".
• →Graduate deliberately, not in a rage. When your work becomes 80% maintenance and 20% new features, the honest move is a code editor with an AI in it. That is not betrayal. That is the tool doing its job and handing you the result.

The uncomfortable comparison

There is a widely shared argument that Lovable costs roughly $2,400 a year against a one-off $3,000 custom build you own outright, and that the subscription therefore loses. It is a tidy piece of arithmetic and it is missing the second year, in which the custom build needs changing and the developer who wrote it is on a different continent charging day rates. It is also missing the version where you spend the $3,000 building the wrong thing very professionally.

But it is not wrong to ask the question, and the fact that people now ask it about Lovable rather than about WordPress hosting is itself the story. Two years ago the entire debate was whether AI builders could produce anything real. Now the debate is about unit economics and migration paths. That is what a market maturing sounds like. It sounds like complaining.

“The forums have stopped asking whether it works. They have started arguing about the bill. That is not decline — that is what winning sounds like from the cheap seats.”

The verdict

Slower: partly true, partly your project getting fat, partly documented incidents the company publishes itself. More expensive: genuinely true in feel and often in fact, driven by unified billing and by the debugging loop rather than by a price list. Everyone leaving: no — $300m to $400m to $500m annualised inside six months, a million new projects a week, and a valuation multiple that fell because revenue outran it.

What is true is that Lovable has become a platform you have to be competent at, rather than a toy that flatters you. The people shouting loudest about cost are, with striking regularity, the people who have not changed a single habit since their first prompt. The meter is not lying to you. It is describing you.

And if that sounds harsh, consider the alternative on offer. Somewhere on the other side of this argument is a stack that turned 30 years old, needs 40-odd plugins to reach parity, hands you a pre-auth RCE every other quarter, and still charges you hosting rent for the privilege. Pay the credit bill. Learn to prompt. Keep the GitHub export warm. It is still, by a distance, the least stupid deal in the room.