We are at an inflection point that will have a huge impact on how we write software: Meta is investing billions of dollars into building an LLM that is capable of writing software. They are doing this to write software at Facebook—but as with React, what happens at Facebook does not stay at Facebook.
Engineers as training data
In March 2026, Meta involuntarily transferred roughly 6,500 engineers and product managers into a new unit called Applied AI (internally: Agent Data Optimization) whose job is to manufacture training data for Meta's models: hard coding problems, the tests that check them, and grades on AI-written code. According to The Pragmatic Engineer, 30–50% of engineers on core product teams were reassigned—people who were building Instagram, ads, and infrastructure the week before.
Meta considers hand-written, novel coding problems from senior engineers so valuable that it pulled thousands of them off product work to produce them. In leaked audio, Zuckerberg reportedly argued that Meta's own engineers are simply smarter than outside contractors—his case for conscription over hiring.
These engineers are expected to deliver two such tasks per week—with a median software engineer compensation of $444,000 these are some of the most expensive—and arguably best—engineers in the industry.
Aggressive pricing, by design
The second piece landed this week. On July 9, Meta released Muse Spark 1.1, an agentic coding model, together with its first serious paid developer API:
Since this is not an open source model, this is I think the first time that we're doing a real serious API. And the pricing is going to be very aggressive and attractive. — Mark Zuckerberg, Bloomberg interview, 2026-07-09
Developers can use the model for free up to a token threshold, and the API is priced at roughly 25% of what OpenAI and Anthropic charge for their top models. This is a deliberate land grab—and, I believe, an admission that there is no money in selling access to LLMs in the long run. The race to the bottom of frontier model pricing has already started—Meta's own launch at 25% of competitor pricing is that race, and cheap models from other labs are pushing in the same direction. Zuckerberg himself calls the margins of the other labs "very extreme"; his stated strategy is to get the technology in front of as many people as possible. That is not the strategy of someone who plans to live off token revenue. Meta's actual business is advertising, and its actual interest in this model is writing Meta's own software cheaper and faster.
The paid API is just temporary—training models and running inference is expensive today, but we have seen that for example Chinese companies are able to produce coding models that can compete with the latest Western models, having only much more limited access to skills and hardware—training and inference costs will eventually be low enough that Meta can afford to open source their frontier coding model.
And running them will be feasible for anyone, not just a few frontier AI companies with access to thousands of Nvidia GPUs.
Zuckerberg can afford to play this game for as long as he wants. He is the shareholder with super-voting rights, which means he decides—by himself—where Meta invests its money. He is on this track now, and he believes the future of Meta is, in part, to reduce the amount of software written by humans at Meta. Meta's internal targets say as much: the Creation organization aims for 65% of engineers to produce more than 75% of their committed code with AI in the first half of 2026.
Which puts the model in exactly the same position React was in: a tool built to solve Facebook's problem, where giving it away costs Meta little and buys them an ecosystem, mindshare, and a labor market pre-trained on their way of working.
We have seen this movie: React
If you think "this is a Facebook-internal problem", let me remind you of React.
React was open-sourced at JSConf US in May 2013 as a better way to build front-end applications. It solved a problem for Facebook—and there are arguably very few companies in the world that have Facebook-level problems with their frontend applications. Yet today React is the predominant technology for building web applications; it has topped the Stack Overflow developer survey as the most-used web framework for years, in recent years by large margins.
To be fair: Meta has an enormous engineering organization. Of its roughly 78,000 employees at the end of 2025, about a third work in engineering—in the order of 25,000 engineers. React is technically a great achievement. It is an innovation. It solves real problems.
But there were, and are, many good alternatives. React won in large part because enough tech leaders were drinking the Facebook Kool-Aid: what Facebook does must be good, therefore it is good enough for us. Add the job-market flywheel—everybody who wants to work at a FAANG company learns React, so everybody hires for React—and you get the de facto standard we all live with today. An ecosystem that exists purely because Facebook willed it into existence.
Facebook changed the way we build frontends, Meta changes the way we build software
Now run the same playbook with an LLM instead of a UI library.
Meta makes the model cheap—25% of the competition—free to start with, and eventually, I believe, open source. That is the React move: outsource the cost of onboarding. Engineers learn the tool before they apply at Meta, and arrive with the skill Meta needs already available. Meanwhile Facebook is still the biggest website on the planet, and what Meta does technically is still seen by many as the epitome of scalable web application development.
So when Meta at some point says: we no longer employ people who write source code by hand—we employ software engineers, but they direct, review, and grade code instead of writing it (and remember, at Facebook that code is written in Hack, their own PHP dialect introduced in 2014, which tells you how comfortable they are with defining their own way of doing things)—then a significant number of companies around the world will copy that model. Not because it is the best way to build software. React was not the best way either. It will be a controversial way, and better alternatives will continue to exist, including simply writing source code by hand.
But the Facebook way of developing software without writing source code will be impossible to ignore for us as software crafters.
Given the trajectory Meta is on—the billions in investment, 6,500 engineers producing training data, a coding model priced to undercut everyone and, eventually, given away—this will lead to a significant shift in how we build software.
React did not become the standard because it was the best option. It became the standard because Facebook made it free, made it huge, and made it a hiring requirement. The same machine is now being pointed at source code writing itself.
If you like it or not.