AI is transforming how software engineers do their jobs. Just don’t call it ‘vibe-coding’

One of the hottest markets in the artificial intelligence industry is selling chatbots that write computer code.

Some call it “vibe-coding” because it encourages an AI coding assistant to do the grunt work as human software developers work through big ideas. Others dislike that term. But there’s no question that these tools are transforming the job experience for many tech workers amid an intense rivalry between leading AI companies to make the best one.

“The essence of it is you’re no longer in the nitty-gritty syntax,” said Cat Wu, project manager of Anthropic’s Claude Code. “You’re not looking at every single line of code. You’re more trying to communicate this higher-level goal of what you want to accomplish.”

Wu added, however, that ”vibe-coding” is not a term she uses. “We definitely want to make it very clear that the responsibility, at the end of the day, is in the hands of the engineers.”

Anthropic launched the latest version of its flagship Claude chatbot on Monday, boasting that Claude Sonnet 4.5 will be the “world’s best” for coding and other complex tasks.

Large language models behind generative AI chatbots like Claude, ChatGPT and Google’s Gemini are capable of many things, from homework help to organizing meal plans, but the “top use case” for most businesses has been in coding and software engineering, said Gartner analyst Philip Walsh.

“That is often the first thing large organizations go after,” Walsh said. “I think there’s broad recognition among these AI model providers that coding is really where they’re getting the most traction.”

And while Walsh said Anthropic’s products are a favorite for software developers, it is hardly the only player in a rapidly growing and consolidating market.

San Francisco and the surrounding Bay Area remain the center of the battle to make the best AI coder, home not just to fierce rivals OpenAI and Anthropic but startups like Anysphere, Cognition and Harness, as well as Microsoft-owned GitHub.

“This is the most competitive space in the industry right now,” said Windsurf CEO Jeff Wang, speaking by video call from the startup’s office in Mountain View, California.

Windsurf’s coding assistant launched less than a year ago, but as its popularity grew, hitting 200,000 users in its first two months, it quickly found itself at the center of a bidding war between tech giants. OpenAI sought to acquire it. Then, Google scooped up Windsurf’s founders and research team, leaving a shell of a company that another AI coding startup, Cognition, acquired in July.

“It’s been a really volatile time at Windsurf,” Wang told employees in a July email as he announced the merger with Cognition, maker of the AI coding assistant Devin. Two months later, the two companies’ integration is “going really well,” Wang told The Associated Press from a conference room called New Kelp City, named for a fictional setting in SpongeBob SquarePants.

Some AI coding assistants automatically finish the code that human programmers are writing, much like the “autocorrect” features that suggest the next lines of an email or text. More advanced tools known as AI agents are given more autonomy to access computer systems and do the work themselves.

Anthropic said its new Claude Sonnet 4.5, on a test before its public release Monday, was able to code autonomously for more than 30 hours on a project for London-based startup iGent.

Anthropic’s first coding assistant was developed largely by accident when the company’s Boris Cherny built an internal toy project and started using it to accelerate his own work. Then the rest of his team adopted it.

“Over time, we realized that it was just virally spreading within Anthropic,” Wu said.

Anthropic, in a consumer usage report earlier this month, said coding is the top use for Claude, with about 39% of its users saying they use the chatbot for coding.

OpenAI, by contrast, says writing is the most common work task for ChatGPT, with coding and self-expression as more “niche” activities on the platform. Even so, OpenAI has sought to catch up, introducing in September a new GPT-5-Codex that it says can work for longer on complex coding tasks.

Among the most coveted customers for big AI model developers are coding startups like Anysphere, maker of the popular coding tool Cursor, which relies heavily on Anthropic’s Claude and recently cemented a partnership with OpenAI.

It was Cursor’s Composer, combined with Anthropic’s Claude Sonnet, that prominent AI researcher Andrej Karpathy was playing with when he coined the phrase “vibe-coding” in February.

“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,” he wrote on X.

It was “getting too good,” he said, so much so that he could speak his instructions and “barely even touch the keyboard” and use it for throwaway weekend projects.

“It’s not really coding – I just see stuff, say stuff, run stuff, and copy-paste stuff, and it mostly works.”

Anthropic shipped Claude Code a few weeks later.

Some platforms, like Sweden-based Lovable, cater to vibe-coders with an approach that encourages anyone to “create apps and websites by chatting with AI.” But most tools are designed for professionals with programming expertise.

The phenomenon has raised fears of job loss in software careers, fueled by comments from tech CEOs who say AI is speeding up software development and making their teams more efficient.

Walsh said Gartner’s position is that AI will not replace software engineers and will actually require more.

“There’s so much software that isn’t created today because we can’t prioritize it,” Walsh said. “So it’s going to drive demand for more software creation, and that’s going to drive demand for highly skilled software engineers who can do it.”

Economists, however, are also beginning to worry that AI is taking jobs that would otherwise have gone young or entry-level workers. In a report last month, researchers at Stanford University found “substantial declines in employment for early-career workers’’ — ages 22-25 — in fields most exposed to AI.

Stanford researchers also found that AI tools by 2024 were able to solve about 72% of coding problems, up from just over 4% a year earlier. It’s likely to have grown even higher since then.

Karpathy didn’t respond to requests for comment. But the idea that non-technical people in an organization can “vibe-code” business-ready software is a misunderstanding of what Karpathy meant when he came up with the term, Walsh said.

“That’s simply not happening. The quality is not there. The robustness is not there. The scalability and security of the code is not there,” Walsh said. “These tools reward highly skilled technical professionals who already know what ‘good’ looks like.”

Wu said she’s told her younger sister, who’s still in college, that software engineering is still a great career and worth studying.

“When I talk with her about this, I tell her AI will make you a lot faster, but it’s still really important to understand the building blocks because the AI doesn’t always make the right decisions,” Wu said. “A lot of times the human intuition is really important.”