The AI code wars are heating up

This is The Stepback, a weekly newsletter breaking down one essential story from the tech world. For more on the AI coding and vibe-coding booms, follow David Pierce. The Stepback arrives in our subscribers’ inboxes at 8AM ET. Opt in for The Stepback here.

Writing code was a killer app for AI even before anyone was really talking about AI. In the spring of 2021, 18 months before the world knew the word “ChatGPT,” Microsoft debuted the very first product of a partnership with a nonprofit called OpenAI: a tool called GitHub Copilot that watched developers as they wrote code and tried to autocomplete snippets and lines for them. It wasn’t all that good, and it was only a “restricted technical preview,” but more than a million developers signed up to try it anyway.

Large language models seemed obviously poised to make software development even simpler and even faster. Most code is relatively structured and straightforward; coding languages are generally extremely well-documented; and a vast amount of code is available online for use in training models (albeit via sometimes dubious means). Unlike so much other information you might get from an LLM, you can also check the quality of code just by trying to run it. At first, a few companies figured, LLMs might be able to make writing code faster by predicting the next word the way Google’s autocomplete might. But pretty soon, they hoped, it might be able to do some of the coding for you. Maybe even all of it.

For so many years, companies around the tech industry had also pursued the idea of “low code” and “no code” software. Rather than offering users endless lists of settings and unparseable menus, the idea was to effectively let people build software themselves. For a long time, this was pretty hacky: you got things like Zapier and Apple Shortcuts, which were effectively super-complex workflow builders; or you got software like Notion and Airtable, which were immensely flexible at the cost of being pretty hard to figure out.

Even in those early days, it was also obvious why AI coding tools might one day be a good business. Developers are expensive; product creation takes a long time. Any tool that might mean companies could hire fewer developers, or help developers be more productive, would surely be an easy pitch to software companies the world over. If the tech ever worked, the products would practically sell themselves. Companies like Cursor and Windsurf raised huge sums of money to try and build companies around AI coding tools, while OpenAI, Google, Anthropic, and others began building new products for developers.

At first, AI coding tools were not to be trusted. For a couple of years, they could maybe complete a few lines of code, but always needed to be checked. In late 2023, Simon Willison, a programmer and blogger, called LLMs “weird coding interns.” He wondered whether these interns would make coders more versatile and powerful than ever, or eventually begin to replace them.

In early 2025, Anthropic released a product called Claude Code that would soon make that question much more urgent for many more people.

In late 2025, Anthropic released a new version of its Claude LLM, called Opus 4.5. According to Anthropic’s benchmarks, it was the best Claude model yet, but didn’t seem to represent some earth-shattering advancement in AI technology. A few weeks later, though, a lot of developers with a few free hours during the holidays began to test the new model in Claude Code, and almost universally seemed to arrive at the same conclusion: it works. Suddenly, the tool you previously had to carefully prompt and carefully review could turn a few sentences into a working prototype. Boris Cherny, the creator of Claude Code, professed to already having AI write 100 percent of his code. “It was just as surprising for me as it was for everyone else,” he told The Verge earlier this year. In a way that seemed impossible for a coding tool, Claude Code went viral.

Claude Code may have captured a lot of the software world’s imagination, but Anthropic’s competition hasn’t been far behind. OpenAI’s Codex, which launched in 2025 a few months after Claude Code, has gotten a series of updates and is also a powerful and popular tool for writing code. Google rolled out a command line interface for its Gemini model and has recently been putting more coding features into its AI Studio app.

Increasingly, AI coding seems like the first truly mainstream AI use case — not to mention the first potentially great AI business. The Claude Code moment coincided with an absolute explosion in revenue for Anthropic; one of OpenAI’s top executives recently told her team to stop doing “side quests” and focus instead on competing with Anthropic and Claude Code. Both OpenAI and Anthropic are reportedly planning to go public this year, which means both companies will need something to show for the billions they’ve raised in capital, and the billions they’ve burned on compute. It seems everyone’s best idea is writing code.

In fairness, it looks like a pretty reasonable guess. Companies around Silicon Valley are suddenly seeing employees compete to use the most tokens, using GPU access as a recruiting tool, and bragging publicly about their AI bills. Nvidia CEO Jensen Huang recently said he’d worry about any highly paid engineer who wasn’t spending $250,000 a year on AI tokens. Even as developers fear AI coding tools might spell the end of their careers and livelihoods, the race is on to embrace them as quickly as possible. One 2025 study found that 98 percent of respondents said they used AI coding tools “several times a week.”

It’s not just developers, either. In February of 2025, Andrej Karpathy, a veteran of the AI industry, coined the term “vibe coding.” “I’m building a project or webapp,” he wrote on X, “but it’s not really coding – I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.”

It’s a shame Karpathy didn’t come up with a catchier phrase, because vibe coding stuck. The name as well as the phenomenon: lots of people who didn’t or even couldn’t write code were suddenly prompting their way to workable software. For many of those people, who might otherwise have made slide decks or Figma mockups, a barely functional prototype was plenty, and these coding tools have proven more than capable of building barely functional prototypes. Vibe coding does come with risks, though, both in terms of the problems bad code can cause and the risks you run by giving these tools access to your computer and your data. It’s one thing to trust the system when you can verify its output, another to do so when you can’t speak its language.

The software developer crisis is only just beginning. Companies around Silicon Valley are laying off employees by the thousands, usually citing AI as the reason. “A significantly smaller team, using the tools we’re building, can do more and do it better,” Block CEO Jack Dorsey wrote in a memo announcing 40 percent of the company was being laid off. “And intelligence tool capabilities are compounding faster every week.” In Block’s case and many others, AI is likely at least in part just a cover for pandemic-era overhiring, but the tech industry is clearly set on AI as a way to enhance productivity — and reduce headcount.

As AI coding tools continue to improve, they might also remake the rest of the software business. Why pay a fortune for someone else’s software when Claude Code could build it for you, exactly the way you want it? Some are calling this the SaaSpocalypse, and predicting a fundamental rethinking of the way we value software. Others think we’re due for a new generation of successful startups, which offer AI-native ways to do everything. Still others think it’s all overblown and Salesforce will be just fine. Whatever the outcome, the software industry, which has grown to such unthinkable heights and valuations, feels to many like it’s suddenly on shaky ground.

On the other end of the developer spectrum are the vibe coders. For most people, even the simplest current AI coding tools are too much. They make you read code; they require Terminal access; they ask a lot of questions hardly anyone should be expected to know how to answer. AI coding still comes with plenty of bugs, big privacy questions, and too many ways in which bad actors are able to exploit them both.

With products like Claude Cowork, Anthropic has begun to see if it can make Claude Code’s technology a little more accessible and less intimidating — you just give it access to a bunch of files on your computer and let it go to work. Products like Perplexity Computer are exploring whether people might give LLMs access to everything on their devices, allowing the AI tools to organize files, answer messages, even buy things on their behalf. The underlying tech is beginning to work, but it’s not at all clear how people are supposed to use it, and whether they’ll even want to.