The AI code wars are heating up

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

Writing code was a killer app for AI even earlier than anybody was actually speaking about AI. In the spring of 2021, 18 months earlier than the world knew the phrase “ChatGPT,” Microsoft debuted the very first product of a partnership with a nonprofit referred to as OpenAI: a software referred to as GitHub Copilot that watched builders as they wrote code and tried to autocomplete snippets and features for them. It wasn’t all that good, and it was solely a “restricted technical preview,” however greater than 1,000,000 builders signed up to strive it anyway.

Large language fashions appeared clearly poised to make software program growth even easier and even sooner. Most code is comparatively structured and easy; coding languages are typically extraordinarily well-documented; and an unlimited quantity of code is offered on-line to be used in coaching fashions (albeit by way of generally doubtful means). Unlike a lot different info you would possibly get from an LLM, you too can test the standard of code simply by making an attempt to run it. At first, a number of corporations figured, LLMs would possibly have the ability to make writing code sooner by predicting the following phrase the way in which Google’s autocomplete would possibly. But fairly quickly, they hoped, it’d have the ability to do among the coding for you. Maybe even all of it.

For so a few years, corporations across the tech business had additionally pursued the concept of “low code” and “no code” software program. Rather than providing customers limitless lists of settings and unparseable menus, the concept was to successfully let individuals construct software program themselves. For a very long time, this was fairly hacky: you bought issues like Zapier and Apple Shortcuts, which have been successfully super-complex workflow builders; otherwise you acquired software program like Notion and Airtable, which have been immensely versatile at the price of being fairly arduous to determine.

Even in these early days, it was additionally apparent why AI coding instruments would possibly sooner or later be enterprise. Developers are costly; product creation takes a very long time. Any software which may imply corporations might rent fewer builders, or assist builders be extra productive, would certainly be a straightforward pitch to software program corporations the world over. If the tech ever labored, the merchandise would virtually promote themselves. Companies like Cursor and Windsurf raised enormous sums of cash to attempt to construct corporations round AI coding instruments, whereas OpenAI, Google, Anthropic, and others started constructing new merchandise for builders.

At first, AI coding instruments have been to not be trusted. For a few years, they may possibly full a number of strains of code, however all the time wanted to be checked. In late 2023, Simon Willison, a programmer and blogger, referred to as LLMs “weird coding interns.” He questioned whether or not these interns would make coders extra versatile and highly effective than ever, or finally start to interchange them.

In early 2025, Anthropic launched a product referred to as Claude Code that might quickly make that query far more pressing for a lot of extra individuals.

In late 2025, Anthropic launched a brand new model of its Claude LLM, referred to as Opus 4.5. According to Anthropic’s benchmarks, it was one of the best Claude mannequin but, however didn’t appear to symbolize some earth-shattering development in AI know-how. Just a few weeks later, although, loads of builders with a number of free hours throughout the holidays started to check the brand new mannequin in Claude Code, and virtually universally appeared to reach on the identical conclusion: it really works. Suddenly, the software you beforehand needed to fastidiously immediate and punctiliously evaluation might flip a number of sentences right into a working prototype. Boris Cherny, the creator of Claude Code, professed to already having AI write one hundred pc of his code. “It was just as surprising for me as it was for everyone else,” he advised The Verge earlier this 12 months. In a approach that appeared inconceivable for a coding software, Claude Code went viral.

Claude Code could have captured loads of the software program world’s creativeness, however Anthropic’s competitors hasn’t been far behind. OpenAI’s Codex, which launched in 2025 a number of months after Claude Code, has gotten a collection of updates and can also be a strong and in style software for writing code. Google rolled out a command line interface for its Gemini mannequin and has just lately been placing extra coding options into its AI Studio app.

Increasingly, AI coding looks as if the primary really mainstream AI use case — to not point out the primary probably nice AI enterprise. The Claude Code second coincided with an absolute explosion in income for Anthropic; one among OpenAI’s prime executives just lately advised her crew to cease doing “side quests” and focus as a substitute on competing with Anthropic and Claude Code. Both OpenAI and Anthropic are reportedly planning to go public this 12 months, which implies each corporations will want one thing to indicate for the billions they’ve raised in capital, and the billions they’ve burned on compute. It appears everybody’s finest thought is writing code.

In equity, it appears to be like like a reasonably affordable guess. Companies round Silicon Valley are all of the sudden seeing staff compete to make use of essentially the most tokens, utilizing GPU entry as a recruiting software, and bragging publicly about their AI payments. Nvidia CEO Jensen Huang just lately mentioned he’d fear about any extremely paid engineer who wasn’t spending $250,000 a 12 months on AI tokens. Even as builders worry AI coding instruments would possibly spell the top of their careers and livelihoods, the race is on to embrace them as shortly as doable. One 2025 research discovered that 98 % of respondents mentioned they used AI coding instruments “several times a week.”

It’s not simply builders, both. In February of 2025, Andrej Karpathy, a veteran of the AI business, coined the time period “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 disgrace Karpathy didn’t come up with a catchier phrase, as a result of vibe coding caught. The title in addition to the phenomenon: tons of people that didn’t and even couldn’t write code have been all of the sudden prompting their approach to workable software program. For lots of these individuals, who would possibly in any other case have made slide decks or Figma mockups, a barely purposeful prototype was lots, and these coding instruments have confirmed greater than able to constructing barely purposeful prototypes. Vibe coding does include dangers, although, each by way of the issues unhealthy code may cause and the dangers you run by giving these instruments entry to your laptop and your information. It’s one factor to belief the system when you’ll be able to confirm its output, one other to take action when you’ll be able to’t converse its language.

The software program developer disaster is simply simply starting. Companies round Silicon Valley are shedding staff by the hundreds, normally citing AI as the explanation. “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 saying 40 % of the corporate was being laid off. “And intelligence tool capabilities are compounding faster every week.” In Block’s case and lots of others, AI is probably going at the least partly only a cowl for pandemic-era overhiring, however the tech business is clearly set on AI as a approach to improve productiveness — and scale back headcount.

As AI coding instruments proceed to enhance, they could additionally remake the remainder of the software program enterprise. Why pay a fortune for another person’s software program when Claude Code might construct it for you, precisely the way in which you need it? Some are calling this the SaaSpocalypse, and predicting a elementary rethinking of the way in which we worth software program. Others assume we’re due for a brand new technology of profitable startups, which provide AI-native methods to do all the things. Still others assume it’s all overblown and Salesforce can be simply nice. Whatever the result, the software program business, which has grown to such unthinkable heights and valuations, feels to many prefer it’s all of the sudden on shaky floor.

On the opposite finish of the developer spectrum are the vibe coders. For most individuals, even the only present AI coding instruments are an excessive amount of. They make you learn code; they require Terminal entry; they ask loads of questions hardly anybody ought to be anticipated to know how you can reply. AI coding nonetheless comes with loads of bugs, huge privateness questions, and too some ways by which unhealthy actors are capable of exploit them each.

With merchandise like Claude Cowork, Anthropic has begun to see if it may make Claude Code’s know-how somewhat extra accessible and fewer intimidating — you simply give it entry to a bunch of recordsdata in your laptop and let it go to work. Products like Perplexity Computer are exploring whether or not individuals would possibly give LLMs entry to all the things on their units, permitting the AI instruments to prepare recordsdata, reply messages, even purchase issues on their behalf. The underlying tech is starting to work, nevertheless it’s by no means clear how individuals are supposed to make use of it, and whether or not they’ll even need to.