# What Is Vibe Coding? The Plain-English Meaning, Plus 12 Real Builds and What Each Cost

> Vibe coding means describing software to an AI in plain English and keeping what works without reading the code. These 12 apps, games and websites from Pat Simmons's tests show what each cost and how long it took.

By Pat Simmons · November 10, 2026 · 11 min read

Canonical: https://www.aiformortals.co/blog/what-is-vibe-coding

Vibe coding means building software by describing what you want to an AI in plain English. The AI writes and runs all the code, and you judge the result by trying it instead of reading the code. Andrej Karpathy named it in an X post on February 2, 2025: "There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists."

Pat Simmons, who says he has "no coding experience", built a full agency website this way in Cursor from one 11-page prompt. It took him "20 to 30 minutes just gathering the prompt and then 10 minutes of waiting". More than a year later, Claude Fable 5 spent a bit over 3 hours building a playable Counter-Strike-style browser game from one prompt and two check-ins from Pat, who said the tokens "would have cost $160 if you were using the API".

The gallery below has 12 more builds from his tests, with what each cost and how long it took. The cheapest ways to try it yourself are Cursor's free plan and Codex on ChatGPT's free plan, and Claude Code starts on Claude's $20-a-month Pro plan.

## What does vibe coding mean, and where did the term come from?

Vibe coding is building an app, game or website by telling an AI what you want, accepting the code it writes without reading it, and steering by how the result looks and works. Karpathy described his own habits in that February 2025 post: "I 'Accept All' always, I don't read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it."

A diff is the list of lines the AI changed, so skipping it means trusting the AI's edits without checking them. Karpathy also set the limits in the same post: "It's not too bad for throwaway weekend projects, but still quite amusing."

The phrase spread well past programmers. Collins Dictionary made "vibe coding" its Word of the Year for 2025, announced November 6, 2025, and defines it as "the use of artificial intelligence prompted by natural language to write computer code".

As a verb, "to vibe code" an app or a site means to build it that way.

## 12 things built by vibe coding, with their cost and time

The 12 builds below cost from $1.00 to $67.87 in AI usage each and took 30 to 155 minutes. Every one came from a single prompt that Pat gave several AI models at once, with no code written by hand and no steering while they worked.

_The 12 builds in the table. You can open each one, with its prompt, among all 79 of Pat's builds._

Each row is one model's version of the build. Where Pat ranked a round blind, without knowing which model made which, the Placed column says where that version finished.

BuildTypeModel and toolPlacedCostTimeTested

Scenes from Moby Dick drawn entirely in codeWebsiteGPT-6 Astra in CodexNot ranked blind$8.1630 minSep 2026
20-second launch film with every sound made in codeVideoOpus 5.5 in Claude CodeNot ranked blind$11.6640 minSep 2026
3D skateboarding gameGameFable 5.1 in Claude CodeNot ranked blind$32.2656 minSep 2026
Photoreal 3D bottle you can spin and pour3DOpus 5.5 in Claude Code1st of 4$29.95114 minOct 2026
Rocket League-style car soccer gameGameOpus 5.5 in Claude Code1st of 4$46.24155 minOct 2026
Copy of the oryzo.ai websiteWebsiteOpus 5.5 in Claude Code1st of 4$67.87129 minOct 2026
Product launch video for a coffee brandVideoGPT-6 Astra in Codex2nd of 4$28.4481 minOct 2026
First-person 3D billiards gameGameDeepSeek V4.1 Flash in Claude Code, through OpenRouter3rd of 5$1.0082 minSep–Oct 2026
Rebuild of the heronaiapp.com websiteWebsiteOpus 5.5 in Claude Code1st of 4$42.5763 minBy Oct 12, 2026
Coffee-ordering iPhone appAppOpus 5.5 in Claude Code1st of 4$12.5099 minBy Oct 12, 2026
3D simulation of a Starship launch and tower catch3DFable 5.1 in Claude Code1st of 4$32.9973 minBy Oct 12, 2026
Ad campaign for a cold brew launchBrandingOpus 5.5 in Claude Code1st of 4$19.2847 minBy Oct 12, 2026

The first eight rows come from what the same build cost on each AI model, which lists all 40 runs from three rounds of Pat's tests. The last four come from the Muse Wonder review's four blind-ranked builds.

The costs are API-equivalent, meaning each build's real token count priced at the model's published rate. Pat ran the builds on monthly subscriptions, so these figures show what the same work would cost if you paid per use.

The iPhone app is a native app built in Xcode and run in Apple's iPhone simulator. The billiards game came from DeepSeek V4.1 Flash, an open-weight model (one anyone can download and run), which finished third of five and was the cheapest of the five at $1.00, against $31.74 for Opus 5.5.

## How do you vibe code?

You describe what you want to a coding agent in plain English, let it build, open the result, and tell it what to change until it works. Then you ask the agent how to put it online. Pat's website build, from his May 2025 video, shows each step.

He started with design words. "I don't know any website design terminology," he said, so he gave ChatGPT full-page screenshots of sites he liked and had it describe their style, which came back as phrases like "futuristic but warm", "soft neon glows" and "glassy UI".

Next he wrote the prompt section by section: the headline, a video carousel, the services, a three-step process and the FAQs, each with a screenshot of a site he wanted it to look like. "This is an 11page prompt," he said once it was done.

He pasted all of it into Cursor, running Claude 3.7 Sonnet Max, and let it build. "I know nothing about coding, so I just hit accept all for everything," he said while it built.

_The first version of the PodFlare site from one prompt, open in Chrome. From Pat's video below, at 12:43._

The first version had the services, the process section with small animations and the FAQs. "It's not perfect, but it's like 75 80% of the way there," Pat said. He then switched to the regular 3.7 Sonnet "so I don't go broke" and fixed one section at a time, pasting in a screenshot of the look he wanted for each.

Getting it live took three steps. Cursor pushed the code to GitHub, a site that stores code online, Cursor wrote a configuration file for Vercel, a web host, and Pat imported the GitHub project into Vercel and pressed deploy. His advice for that last step: if Vercel shows an error code, copy it and paste it back into Cursor.

You can start smaller. In our test of Claude Code, one plain-English paragraph about a freelancer's unpaid invoices produced a working app in 90 seconds, and the guide walks you through how to build your first small app with Claude Code.

## Do you need to know how to code to vibe code?

No. Pat built his website and his browser game without writing code, and said so on camera: "I have no coding experience." Partway through the website build he added, "I'll never learn how to code. And I guess I probably don't need to anymore."

What Pat did need was a clear picture of what he wanted and the patience to look at each version and say what was wrong with it. He borrowed the missing vocabulary from ChatGPT, and when he forgot how to put the site online he asked Cursor, "how do I push this to GitHub", and followed its steps.

The same habit worked in our Claude Code test. The first prompt said "I don't know how to code" outright, and the fix for the one thing it got wrong was a plain description of the problem.

## Which AI tools can you vibe code with?

The tools in Pat's builds are coding agents: Cursor, Claude Code and Codex. Each one takes your request in plain English, writes the files and runs them. Pat ran both of his single-prompt builds in Cursor, and his recent model tests ran in Claude Code for the Claude models and in Codex for OpenAI's GPT-6 Astra.

Cursor is an app for building software with an AI agent built in, and Pat picked the model inside it for each build. Claude Code is Anthropic's coding agent and comes with Claude's Pro, Max and Ultra plans. Codex is OpenAI's coding agent and comes with every ChatGPT plan, including Free.

Results also depend on the model inside the tool. In Pat's tests, Opus 5.5 came first in 16 of the 21 apps, games, websites and videos he had models build from identical prompts, 18 of them judged blind, and the full results show which AI model builds best, from 21 blind-judged builds.

If you're choosing between the two tools Pat used most, see Cursor vs Claude Code, by job and price.

## How much does vibe coding cost?

You can start for $0 on Cursor's Free plan or ChatGPT's Free plan, and Claude Code starts at $20 a month. What a single build costs depends on how long the AI works: the 12 builds above came to between $1.00 and $67.87 each at API prices.

ToolFree planPaid plansWhat to know

Claude CodeNot includedPro $20 a month ($17 billed yearly), Max $100, Ultra $200Comes with Pro, Max and Ultra
Cursor$0, with limited agent requestsPlus $8, Premium $30, Super $100, Ultra $200 a monthPlus runs Grok models only; third-party models (from other AI labs) are billed separately at their API list price on Premium and up
Codex$0, for quick coding tasksGo $8, Plus $20, Pro from $100 a monthFree and Go run GPT-6 Luna in the desktop app; Plus adds GPT-6.1 Sol

Prices from Claude's plans page, Cursor's pricing page and OpenAI's Codex pricing page, checked November 10, 2026.

Big builds use a lot of tokens, the small chunks of text an AI reads and writes and bills by, and there's more on what tokens are and why a big build costs more. On Pat's Counter-Strike-style game, Fable 5 used 270K input and 887K output tokens, about $160 at API prices.

_The usage screen Pat opened in Cursor's terminal after the game build. The Fable 5 line shows $160.79, smaller Haiku 4.5 and Opus 4.8 lines bring the session to $180.54, and the build ran 3 hours 18 minutes. From Pat's Fable 5 game video at 3:58._

Pat built the game on a subscription, so he didn't pay that $160 directly, and his usage screen showed 6% of the plan's weekly limit used.

The prices in Pat's website video are from May 2025: Cursor showed "5 cents per request" for Claude 3.7 Sonnet Max then, which is why he dropped to the regular model for his edits.

## What goes wrong when you vibe code

In Pat's builds the problems were unfinished first versions, trouble getting a site online, uneven results, long and costly runs, and code nobody read. He called his first website "75 80% of the way there" and "a little flat", then spent the rest of the video fixing it one section at a time. "You can spend your entire weekend doing that like I just did," he said.

Pat said putting a site online is "typically where people have the most questions". His tip is to paste any error code from the host back into the AI, since the package.json file, which tells Vercel what's in the project, sometimes "doesn't do everything correctly".

_Dust Protocol, the Counter-Strike-style game Fable 5 built in a bit over 3 hours, during Pat's first round. From the same video at 4:25._

Results can be uneven. Pat said of the Fable 5 game that "the actual gameplay is insane. Like this is the first time it feels like that I'm I'm actually playing a video game", but also that "the graphics could be much better... they're just kind of like little robots". He also checked in twice during the build, once to ask for an update and once to ask which model its helper agents were running on, so the game came from one prompt plus two check-ins.

Long runs get expensive. In Pat's tests GPT-6 Astra worked 499 minutes and ran up $357.84 copying the oryzo.ai website, and Fable 5.1 spent $592.37 on a shot-for-shot rebuild of a launch video. Usage limits interrupt too: Claude's limit ran out during the Moby Dick build, and it waited about 30 minutes for a reset.

Some outputs aren't usable at all. In the cold brew ad-campaign test, Pat said none of the four models' launch films was usable, though he still ranked Opus 5.5's campaign first.

And because you don't read the code, you can miss how it works. The invoice app in our Claude Code test kept its list inside the web browser, so clearing your browsing data would erase it, and only Claude Code's own warning made that clear. Karpathy flagged the same trade-off in his original post: "Sometimes the LLMs can't fix a bug so I just work around it or ask for random changes until it goes away."

## Questions people ask about vibe coding

Is vibe coding free?
You can try it free on Cursor's Free plan or with Codex on ChatGPT's Free plan, both with limited use. Claude Code isn't on Claude's Free plan and starts with Pro at $20 a month.

How long does it take to vibe code something?
The 12 builds in the gallery took 30 to 155 minutes of AI work each. Pat's first website took about 30 to 40 minutes including writing the prompt, and his Fable 5 game took a bit over 3 hours.

Can you vibe code a phone app?
Yes. Opus 5.5 built a native coffee-ordering iPhone app in Claude Code for $12.50 in 99 minutes, which ran in Apple's iPhone simulator.

Can you vibe code a game?
Yes. A 3D skateboarding game, a car soccer game and a billiards game in the gallery each came from one prompt, and Dust Protocol came from one prompt plus Pat's two check-ins.

## Update November 10, 2026

This page launched on November 10, 2026 with 12 builds from Pat's tests run in September and October 2026 and plan prices checked that day.

## Sources

- Andrej Karpathy: the "vibe coding" post on X, February 2, 2025

- Collins Dictionary: Word of the Year 2025

- Anthropic: Claude plans, checked November 10, 2026

- Cursor: pricing, checked November 10, 2026

- OpenAI: Codex pricing, checked November 10, 2026

- Pat Simmons: Fable 5 Is BACK And It Just One Shotted THIS (usage screen at 3:58)