Where the term came from
Andrej Karpathy — an OpenAI co-founder who later led AI at Tesla — gave the practice its name in a post on X on 2 February 2025. He was describing his own weekend habit rather than proposing a method: talk to the model, often by voice, accept every change it suggests without looking at the diff, and when something breaks, paste the error message back and let the model sort it out. He was candid that the code grew past the point where he understood it, and that this was fine for projects he meant to throw away.
The phrase caught on at once. Within weeks Merriam-Webster had listed it among its slang and trending words, and in November 2025 Collins named it Word of the Year. Along the way its meaning stretched: many people now say vibe coding for any programming done with an AI assistant, which is not what Karpathy described and not what the rest of this page means.
Vibe coding versus AI-assisted programming
The useful line runs through review, not through tools. Simon Willison, who has written more than most about building with language models, put the test simply in March 2025: if an LLM wrote your code but you reviewed it, tested it and could explain every line to a colleague, you were not vibe coding — you were programming with a fast typist. Vibe coding is the version where nobody reads the code at all.
The distinction matters because every risk discussed below attaches to the not-reading. The same coding agent can produce a careful, reviewed change or a pile of unexamined files; what differs is whether a human takes responsibility for what ships. In practice most professional use sits in between — reading the risky parts, skimming the rest — and it is worth being honest about which mode a given project is in.
What vibe coding costs
There are two ways to pay, and they fail differently.
- A flat plan. Most vibe coding happens inside subscriptions: Claude Pro at $20 a month or Claude Max at $100 and $200, ChatGPT Plus at $20, plus the plans of app builders and AI code editors. The price is fixed; what runs out is the allowance. A long session that keeps asking the model to fix its own bugs empties a usage limit long before the month ends, and then you wait for the reset or upgrade.
- An API key. Coding agents such as Claude Code and Codex can also bill per token. Then nothing stops the session except you: every turn resends the conversation and the files the agent has read, and every failed fix is paid for twice — once to write it, once to diagnose it. See how each agent's routes compare in our AI agents section.
There is a third, easier-to-miss bill: the paid services the finished app calls. Willison's warning is worth repeating — a vibe-coded app wired to a usage-billed API, with no spending cap set, can run up charges nobody budgeted, because nobody read the loop that calls it.
Why vibe-coding sessions burn through limits
A vibe-coding session is an agentic loop with a human who has opted out of the diagnosis. When you paste an error instead of reading the stack trace, the model has to rediscover the problem: open the files again, rerun commands, reason about causes. Each of those steps is a model call that carries the whole conversation so far, so the context — and the token count — grows with every round of “still broken”.
A few habits cut the burn without giving up the speed:
- One feature per session. Start fresh when the topic changes, so old file contents stop riding along in the context window.
- Commit working states. Rolling back to the last good version is free; asking the model to undo its last five attempts is not.
- Keep standing instructions short. Project rules files are resent on every turn.
- Describe the failure, not just the symptom. One sentence about what you expected saves several exploratory turns.
Where it breaks
The criticisms that have stuck are concrete. On security, a May 2025 investigation of apps built with one popular app builder reported that 170 of 1,645 exposed users' personal data — the kind of flaw a review would catch and a demo would not. On maintainability, GitClear's analysis of code changes from 2021 to 2024 reported refactoring falling and duplicated code rising as AI assistants spread — what you get when nobody consolidates what the model writes. And on speed itself, METR's randomized trial in July 2025 found experienced developers working on their own projects were slower with AI tools while believing they were faster.
None of this says the practice is useless. It says the costs arrive later and land on whoever maintains, secures or pays for the result.
When it is the right tool
Vibe coding fits work where a wrong answer is cheap and visible: a personal script, a prototype to test an idea, a one-off data cleanup, a tool only you will use, or learning what a model can and cannot build. It fits badly wherever software touches other people's data, money or safety, or has to be maintained by someone else. The practical rule from people who use it well: vibe code to find out what you want, then read, test and own the version you ship — or have someone experienced review it before anyone else depends on it.