OpenAI has just unleashed GPT-5.3-Codex-Spark on the world—a model that writes code at over a thousand tokens per second. And that’s not all. It runs on specialized chips from Cerebras that contain four trillion transistors. On February 12, the tech world woke up to a new era. OpenAI announced a partnership with Cerebras worth over $10 billion and introduced Codex-Spark—a smaller but lightning-fast version of its coding model. But this move wasn’t just another product announcement. It was a signal that programming is changing faster than we expected.
Breathtaking speed
To understand what Codex-Spark means, we need to go back to how AI models work. Traditional models run on GPU chips from Nvidia—they’re great, but they have their limitations. Cerebras came up with something different: the Wafer Scale Engine 3. The chip is so large that it occupies an entire silicon wafer. It isn’t a small square, but a massive slab packed with computing power.
The result? Codex-Spark can generate responses almost instantly. OpenAI reduced latency by 80%, cut time to first token in half, and reduced per-token overhead by 30%. This allows developers to collaborate with AI in real time—to interrupt it, redirect it, and adjust the course of its work as if it were a live colleague.
Programmers no longer write code by hand
Then came the twist. Within a week of Codex-Spark’s launch, developers around the world began sharing their stories. Boris Cherny, head of Claude Code at Anthropic (an OpenAI competitor), admitted that he hadn’t written any code in two months. Gustav Söderström, co-CEO of Spotify, said during an earnings call that the company’s best developers hadn’t written a single line since December. Instead, they send instructions to AI through Slack—perhaps from their phones on the way to work—and by the time they arrive at the office, completed features are ready for deployment. Thanks to this, Spotify released more than 50 new features in 2025.
But the wildest part of the story? AI models are now writing themselves. OpenAI admitted that GPT-5.3-Codex "is our first model that was instrumental in creating itself." Anthropic said that AI writes 70 to 90% of its code. Even Claude Code—a tool for writing code—was 90% written by itself.
Programmers’ existential crisis
Matt Shumer, CEO of OthersideAI, wrote an essay that spread across the internet like an avalanche. He claimed that AI models can now handle the entire development cycle autonomously—they write tens of thousands of lines of code, test it, fix bugs, and repeat the process until they are satisfied. Developers supposedly just describe what they want and walk away. Shumer warned that AI could disrupt the labor market more than the COVID pandemic.
Reactions were mixed. Alexis Ohanian, co-founder of Reddit, agreed. Gary Marcus, a professor at NYU, called it "weaponized propaganda" and noted that Shumer had provided no data. Fortune’s Jeremy Kahn argued that coding has unique characteristics—such as automated testing—that make it easier to automate than other fields.
But for many engineers, Shumer’s words were not a warning, but a description of reality. Developers at leading tech companies have not been writing code line by line for months. They have become directors of AI systems that do the writing for them. The skill has shifted from writing code to designing solutions and guiding AI tools.
The dark side of speed
And that’s not all. Steve Yegge, a veteran engineer, warned about burnout. In a blog post, he described how he falls asleep after long coding sessions and how his colleagues are considering installing reclining chairs in the office. He said the addictive nature of AI tools is pushing developers toward unsustainable workloads. "With a 10x acceleration, when you give an engineer Claude Code, their work creates the value of nine additional engineers," Yegge wrote. "But building things with AI requires a great deal of human energy."
A new chapter in programming
OpenAI sees Codex-Spark as the first step toward a dual-mode workflow: long-term reasoning and execution on one side, and real-time collaboration for rapid iteration on the other. Over time, these modes will merge—Codex will keep you in a tight interactive loop while delegating longer-term work to background sub-agents. Cerebras and OpenAI are now working to expand data center capacity and deploy larger models. For now, Codex-Spark is available only to ChatGPT Pro users as a research preview, but access will expand.
So, is traditional coding dead? Maybe not entirely. But it is certainly dying in the form we knew. Developers are not turning into dinosaurs—they are becoming conductors of an orchestra of AI tools. And those who adapt the fastest will have an advantage. As Sam Altman said in a tweet before the announcement: "We have something special for Codex users on the Pro plan. It brings me joy." And perhaps that very joy—or fear—is what we are all feeling now. Because the future of programming is no longer the future. It is here. Now.
