Anthropic published a report on Thursday that should have been front-page news everywhere. It wasn't, because the company wrapped a genuinely alarming disclosure inside a policy proposal it knows is impossible.
The disclosure: Claude now writes more than 80% of the code merged into Anthropic's codebase. Eighteen months ago that figure was in low single digits. On the hardest coding tasks, Claude's success rate jumped 50 percentage points in six months. An internal benchmark measuring how much faster each new model can make training code run went from three times faster with Claude Opus 4 to 52 times faster with the unreleased Mythos Preview.
The AI is building the AI. Not metaphorically. Literally.
A proposal that cancels itself
Anthropic's response to its own data was to call for a globally coordinated option to slow or pause frontier AI development. The company said a pause would only be effective if multiple labs across multiple countries, including the US and China, agreed to stop simultaneously under verifiable rules.
That framework does not exist. No treaty mechanism covers it. No inspection regime could enforce it. China has no incentive to participate. OpenAI is weeks from an IPO. Google is spending $190bn on AI infrastructure this year. Nobody is stopping.
Anthropic knows this. The report says so explicitly: a unilateral pause changes who leads without achieving any safety benefit.
So the company has published a paper arguing that the world needs a fire extinguisher while continuing to pour petrol. The intellectual honesty is real. The practical impact is zero.
Recursive self-improvement problem
The concept Anthropic is warning about, recursive self-improvement, is the scenario AI safety researchers have flagged for a decade. An AI system that can design and build its own successor without meaningful human involvement enters a feedback loop where each generation is more capable than the last, and the gap between what the system can do and what humans can understand widens with every cycle.
Anthropic says this has not happened yet. It also says it could happen within two years. Co-founder Jack Clark warned that small misalignments in current models could compound across generations, becoming harder to detect with each iteration.
The 80% code figure is not recursive self-improvement. But it is the on-ramp. A model that writes most of the code that builds the next version of itself is one architectural step away from closing the loop entirely.
Brand play
Anthropic has always positioned itself as the responsible lab. The one that cares about safety. The one that would rather be right than first.
That positioning is commercially valuable. It attracts safety-conscious enterprise customers. It gives regulators a company to point to when they need an industry partner. It differentiates Anthropic from OpenAI, whose approach to safety has been, charitably, more flexible.
The "When AI Builds Itself" report extends that brand. It says: we see the risks more clearly than anyone, and we are telling you about them. If something goes wrong, remember that we warned you.
The uncomfortable truth
The report is not wrong. The data is genuine. The risks are real. The call for coordination is logical. And none of it will result in anyone slowing down, including Anthropic.
The company filed a confidential S-1 with the SEC four days ago. It is valued at $965bn. Its revenue run rate is $47bn. It is not pausing.
Anthropic is asking the world for permission to stop while running faster than it ever has. The sincerity of the concern and the impossibility of the solution are not contradictory. They are the point.
The company has told you the building might be on fire. It has also told you it will not leave until everyone else does. Make of that what you will.
