Last week, OpenAI and Anthropic dropped their respective coding models—GPT-5.3-Codex and Claude Opus 4.6—both of which represented significant leaps in AI coding capabilities. GPT-5.3-Codex showed markedly higher performance on coding benchmarks than earlier models, while Opus 4.6 introduced a feature that lets users deploy autonomous AI agent teams that can tackle different aspects of complex projects simultaneously. Both models can write, test, and debug code with minimal human intervention—even iterating on their own work and refining features before presenting results to developers.

The releases—especially GPT-5.3-Codex—sparked something of an online existential crisis among software engineers. At the heart of it was a viral essay written by Matt Shumer, CEO of OthersideAI. Shumer said that “something clicked” following the model releases and described AI models now handling the entire development cycle autonomously—writing tens of thousands of lines of code, opening applications, testing features, and iterating until satisfied, with developers simply describing desired outcomes and walking away. He proposed that the advances meant that AI could disrupt jobs more severely than the COVID-19 pandemic.

The essay drew mixed reactions. Some tech leaders agreed, including Reddit co-founder Alexis Ohanian, but others, including NYU professor Gary Marcus, criticized it as “weaponized hype.” (Marcus noted that Shumer provided no data supporting claims that AI can write complex apps without errors.) Fortune’s Jeremy Kahn also argued that it was coding’s unique characteristics—like automated testing—that made it easier to fully automate, while the automation of other knowledge-work fields may be more elusive.

Software engineers as early adopters

For many engineers, some of Shumer’s warnings just reflect their current reality. Many engineers say they have stopped coding entirely, instead relying on AI to write code at their direction.

While the new releases do represent meaningful improvements, developers also said the industry has been undergoing a slow transformation over the past year as models became capable enough to handle increasingly complex tasks autonomously. While developers at leading tech companies have largely stopped writing code line-by-line, they haven’t stopped building software—they’ve become directors of AI systems that do the typing for them. The skill has transformed from writing code to architecting solutions and guiding AI tools. The new models, some argue, mainly “burst the bubble” around AI coding by making people outside coding aware of a trend engineers have been experiencing for months.

During its earnings call this week, Spotify co-CEO Gustav Söderström said the company’s best developers “have not written a single line of code since December.” The streaming giant’s internal system uses Claude Code for remote deployment, allowing engineers to instruct AI to fix bugs or add features via Slack on their phones during their commute, then merge completed work to production before reaching the office. Söderström said Spotify shipped over 50 new features in 2025 using these workflows.

Even within Anthropic, engineers are heavily relying on their own tools to write new code. Boris Cherny, head of Claude Code, said earlier this month that he hasn’t written code in over two months. Anthropic previously told Fortune that 70% to 90% of the company’s code is now AI-generated.

The models themselves have also reached a recursive milestone: They’re now materially helping to build more advanced iterations of themselves. OpenAI said GPT-5.3-Codex “is our first model that was instrumental in creating itself,” a significant shift in how AI development works. Similarly, Anthropic’s Cherny said his team built Claude Cowork—a non-technical version of Claude Code for file management—in approximately a week and a half, largely using Claude Code itself. Even for Claude Code, Cherny said about 90% of its own code is now written by Claude Code.

Despite the productivity gains, some developers are also warning that the new tools could result in burnout. Steve Yegge, a veteran engineer, said that AI tools were draining developers through overwork.

In a widely shared blogpost, Yegge described falling asleep suddenly after long coding sessions and colleagues considering installing nap pods at their office. The addictive nature of AI coding tools, he argues, is pushing developers to take on unsustainable workloads. “With a 10x boost, if you give an engineer Claude Code, then once they’re fluent, their work stream will produce nine additional engineers’ worth of value,” he wrote. But “building things with AI takes a lot of human energy.”