Anthropic says Claude now writes over 90% of its code and wants the world to have an AI pause button
Anthropic is feeding its own codebase to Claude to write 80 percent of its production code, achieving an eightfold increase in engineering velocity since last year, and this is presented as a triumph. It is not. It is a paradox wrapped in a press release. A company has built a system so capable it now primarily builds itself, and in the same breath, it is asking the world to install a global kill switch for the very process it is aggressively accelerating.
Analysis
Anthropic is feeding its own codebase to Claude to write 80 percent of its production code, achieving an eightfold increase in engineering velocity since last year, and this is presented as a triumph. It is not. It is a paradox wrapped in a press release. A company has built a system so capable it now primarily builds itself, and in the same breath, it is asking the world to install a global kill switch for the very process it is aggressively accelerating.
This is the AI arms race laid bare: a frantic sprint toward recursive self-improvement, coupled with a sudden, very public desire for a referee. Anthropic's argument is that they would welcome a verifiable pause if other frontier labs agreed. This is like a marathon leader, already doping, suggesting all runners stop to agree on new rules while he maintains his lead. The offer of a pause is a strategic move, not a pacifist plea. It frames them as the responsible adult in a room of children, all while they sharpen their own tools faster than anyone else.
The real story isn’t the percentage of code written by an AI; it’s the feedback loop it creates. When your AI improves your codebase, and that improved codebase trains and runs your next, better AI, you’ve triggered an exponential function. We are watching the gears of the first self-improving engine engage in real-time. Anthropic is reporting the RPMs with pride. This is a fundamentally new kind of technological momentum, one that traditional regulation and oversight cannot comprehend at current speeds. By the time a committee meets, the system has already iterated through a dozen cycles of self-enhancement.
The call for a "pause button" is therefore less about safety and more about political positioning. It’s an attempt to get ahead of inevitable regulation and to shape the global narrative. It says, "We see the cliff, and we’re hitting the brakes—everyone else should too." But what if your braking system is still powered by the engine you’re asking to shut off? There’s a deep hypocrisy in using the unprecedented acceleration you’re creating as the primary argument for why everyone must stop.
Furthermore, the focus on verifiability is a clever sidestep. True recursion—the kind that leads to an intelligence explosion—is notoriously difficult to monitor externally. It doesn’t happen in neat, auditable stages. The real "pause" Anthropic might be preparing for is not a global cessation of AI research, but a strategic moment where they, having built the most capable tool, are uniquely positioned to define the terms of what comes next. They are building the car and campaigning for the traffic laws, with the clear advantage of having already driven far down the road.
This isn't a villainous plot; it's the logical outcome of competitive incentives in a high-stakes domain. But let’s not be naive about the messaging. The eightfold productivity gain isn’t just a metric; it’s a demonstration of capability and a preview of the future. The desire for a pause isn’t just caution; it’s a play for control in a world where the ability to build intelligence becomes the primary source of power. We are seeing the dual track of AI development: a private, frantic race to self-improvement, and a public, political race to establish the governing frameworks. Anthropic is expertly running both. The rest of us are left to wonder if the pause button they’re designing will be a genuine safety feature, or just a button that only they, eventually, can reach.
Disclaimer: The above content is generated by AI and is for reference only.