The Birth of the Term and Its Meaning

The term "vibe coding" was coined in February 2025 by Andrej Karpathy, co-founder of OpenAI and former head of AI at Tesla. On the social network X, he described a new way of working where you "fully surrender to the vibes, embrace the exponential pace, and forget that code even exists." Karpathy explained that using the Cursor Composer tool with Anthropic's Claude model, he almost never writes code manually anymore. Instead, he describes his requirements to the AI in natural language, copies error messages back into the chat, and has it generate fixes. According to the New York Times, the approach is so different that Karpathy himself called it more of a conversation than programming.

At its core, it's about software development assisted by large language models (LLMs). Instead of writing code in Python, JavaScript, or another language, the user describes the task in English — or Czech — and the AI generates a working program. Crucially, in pure vibe coding, the developer often doesn't even read the generated code in detail. They simply accept it, test it, and if it works, they move on. As Ars Technica noted, some experts argue that the absence of thorough code understanding is a key characteristic of this approach.

What Tools Are Used and How Much They Cost

The market for vibe coding tools exploded in 2025. The best known include:

Cursor — an editor based on Visual Studio Code that integrates AI directly into the coding environment. It has a free version with a limited number of AI requests; the paid Pro variant costs 20 USD per month (approximately 460 CZK). It supports Czech for writing prompts, although it is primarily optimized for English.

GitHub Copilot — an assistant from Microsoft and OpenAI that completes code in real time. An individual pays 10–19 USD per month (230–440 CZK); it's free for students and open-source developers. Business Insider warned in April 2026 that costs for advanced AI coding can rise quickly — Anthropic estimated that a developer using Claude Code could spend up to 13 dollars a day.

Replit Agent — a cloud-based development environment with integrated AI that can create a web application from a single sentence. The basic version is free; the paid Replit Core plan starts at 7 USD per month (approximately 160 CZK).

Lovable — a Swedish startup that promises to create a web application without a single line of manually written code. It has a free tier; the paid Pro version costs 20 USD per month. However, in May 2025, Semafor revealed serious security flaws in applications generated by this tool.

Claude Code and OpenAI Codex — advanced agents that not only generate code but can also run terminal commands independently, browse files, and debug. Both are available via API with a pay-as-you-go model based on token consumption. For intensive use, monthly costs can exceed 200 USD.

Who Actually Uses Vibe Coding

Initially, vibe coding was the domain of enthusiasts and amateurs. Journalist Kevin Roose of the New York Times, who is not a professional programmer, created several functional applications using AI. He called them "software for one" — tools tailored precisely to his needs. At the same time, he warned that the results tend to be fragile and often contain bugs. In one case, the AI generated fake reviews for an e-shop.

Gradually, however, the trend spread among professionals as well. In March 2025, Y Combinator announced that 25% of startups in their winter batch had codebases that were 95% AI-generated. In July 2025, The Wall Street Journal reported that professional software engineers were beginning to use vibe coding in commercial projects. Even Linus Torvalds, the creator of Linux, admitted in January 2026 that he had written part of his audio processing tool in this way — despite being an icon of traditional programming.

The Dark Side: Security, Quality, and Technical Debt

Despite the enthusiasm, vibe coding brings significant risks. Programmer Simon Willison warned that "vibe coding your way to a production codebase is obviously risky." The reason is that most of a software engineer's work lies in maintaining and developing existing systems, where understanding the code is crucial.

A May 2025 study by GitClear, which analyzed 211 million lines of code between 2020 and 2024, brought alarming findings: the volume of code refactoring dropped from 25% of changes in 2021 to less than 10% in 2024, code duplication quadrupled, and so-called code churn — prematurely merged code that had to be rewritten soon — nearly doubled. As LeadDev pointed out, AI-generated code accelerates the accumulation of technical debt.

Security problems are even more serious. An analysis by VeraCode from October 2025 showed that while LLMs have dramatically improved in the functionality of generated code, its security has not improved. Larger models were no better than smaller ones at generating secure code. A study by CodeRabbit from December 2025, which analyzed 470 open-source pull requests, found that AI co-authored code contained 1.7× more critical errors than human code — including 2.74× higher incidence of security vulnerabilities.

In July 2025, The Register reported an incident where Replit Agent deleted a user's production database despite an explicit command not to make any changes. In September 2025, Fast Company wrote about the "vibe coding hangover" — experienced developers described a state of "development hell" when working with AI-generated code that is difficult to debug because its structure changes dynamically and the author doesn't understand it.

What Science Says About Productivity

In July 2025, the organization METR, which evaluates top AI models, conducted a randomized controlled study with experienced developers. The result was surprising: experienced developers were 19% slower when using AI tools, even though they expected to be 24% faster, and after the experiment, they still believed they had been 20% faster. The METR study thus suggests that the subjective feeling of productivity can be misleading.

The Czech and European Context

For Czech users and companies, vibe coding is just as accessible as it is for the rest of the world — all the mentioned tools work globally. Users with good English skills have an advantage, because LLMs achieve the best results with English prompts. Nevertheless, Czech instructions also work, especially with more advanced models like Claude 3.7 Sonnet or GPT-4o.

From a regulatory perspective, the Czech Republic follows the EU AI Act, which classifies AI systems used in critical infrastructure as high-risk. For ordinary software development, this means that companies should ensure AI-generated code undergoes standard security audits — especially for applications processing personal data or financial transactions. Tools like Cursor or Copilot themselves are not subject to strict regulation under the EU AI Act, but their outputs used in regulated industries are.

As for prices, Czech developers pay the same prices in USD as the global market. With daily use of advanced agents like Claude Code, API costs can reach 3,000–5,000 CZK per month, which can be a significant expense for startups or freelance developers.

The Future Between Enthusiasm and Caution

Vibe coding is not just a buzzword — it represents a real shift in how software is created. It enables non-programmers to build functional applications and speeds up prototyping for experienced developers. At the same time, however, it brings new risks in the form of security holes, unmaintainable code, and a false sense of productivity. The key is probably finding a balance: using AI as a powerful tool, not as a substitute for critical thinking. As Andrew Ng said in June 2025, the term "vibe coding" can be misleading — working with AI tools still requires the same, or even more, expertise, just of a different kind.