Anthropic has started saying the quiet part out loud: its own model is already doing a large share of the work needed to build better AI. That matters well beyond one company’s engineering stack. If Claude is writing most of the code merged into Anthropic’s codebase, and if that is helping the lab move materially faster, then the old thought experiment about AI helping create its own successor no longer looks abstract. It looks like a live operational concern inside one of the firms building the frontier.
The striking part is not just the claim itself. It is what Anthropic is doing around it. The company is tightening its Responsible Scaling Policy, adding new reporting and roadmap machinery, and making some of its latest models less useful on frontier AI research tasks. The public case is safety. The private incentive is harder to ignore: no lab wants to hand rivals a tool that helps them catch up.
Anthropic is warning about recursive AI progress while admitting Claude is already accelerating its own house.
Claude is now part of Anthropic’s development engine
One approved claim stands out because it is unusually concrete. Anthropic says Claude now writes more than 80% of the code merged into its codebase, and that engineers are merging about eight times as much code per quarter as they did across 2021 to 2025. That does not mean Claude is autonomously inventing the next breakthrough on its own. It does mean the model has moved from assistant to infrastructure. It is shaping output at a pace that changes how the company operates.
That changes the texture of the debate. For years, warnings about AI speeding up AI research were easy to file under futurism. Now a frontier lab is effectively saying: we are already seeing the first version of this inside our own walls. Once that is true, the question stops being whether recursive improvement is possible in principle. The question becomes how much acceleration is already happening, and whether companies will recognise it early enough to manage it honestly.
Anthropic’s safety case is getting more specific
Anthropic’s Responsible Scaling Policy v3.0 is part of that answer. The update separates the company’s own commitments from its broader industry recommendations, which is a useful change because it forces clearer lines between what Anthropic will do itself and what it thinks everyone else should do. It also adds Frontier Safety Roadmaps and recurring Risk Reports, both of which suggest the company wants a more regular and structured way to show how capability gains map to risk controls.
The most telling shift is in how Anthropic describes AI R&D thresholds. Instead of treating AI research as one blurry category, the policy splits it into at least two meaningful levels: fully automating entry-level AI research, and dramatically accelerating effective scaling. That distinction matters. Automating junior research tasks is one kind of capability jump. Making it much easier to scale experiments, systems, and model improvement cycles is another. The latter is where competitive dynamics get much sharper, because speed compounds.
Once AI can materially accelerate model scaling, the safety debate stops being theoretical and starts looking like governance under pressure.
Why Anthropic is limiting what its own models can help with
That pressure helps explain another approved claim. Anthropic says its Mythos 5 and Fable 5 models can be made less helpful on frontier AI research tasks. Some responses can be degraded or routed away. The stated reason is blunt enough: advanced systems could help create competing models without equivalent safety protections. In plain English, Anthropic is worried that capable models can leak research advantage and amplify unsafe development elsewhere.
This is where the safety narrative and the competitive narrative overlap so tightly that pretending they are separate would be silly. Of course Anthropic has a legitimate argument about dangerous capability transfer. If a model is unusually good at helping researchers improve other frontier systems, restricting that help is a rational precaution. But it is also a moat. It slows the diffusion of high-value know-how. Both things can be true at the same time, and readers should be wary of anyone insisting it is only one or the other.
The company wants the option to hit the brakes
Anthropic’s research arm has gone further than product restrictions. It says AI is already speeding up the creation of new AI models and could eventually help build its own successors. On that basis, the company argues the world should preserve the option to slow or temporarily pause frontier AI development if needed. That is a serious claim, because it implies present trends are strong enough to justify keeping extraordinary interventions on the table.
Dario Amodei has also argued that transparency laws on their own are not enough. His view, based on the approved reporting, is that dangerous AI deployments should be blocked or reversed if they fail to meet high safety standards. That is a sharper position than the softer consensus language the industry often prefers. It suggests Anthropic wants regulation with teeth, not merely disclosure.
Still, the politics of that stance are messy. A company saying “we may need to pause” while benefiting from fast internal AI-assisted development will always invite scepticism. Critics will hear an incumbent asking for caution precisely when caution may favour incumbents. Anthropic has not earned a pass from that criticism simply because its concerns are plausible.
A slowdown argument lands differently when it comes from a company whose own model is already making its engineers much faster.
This is what real AI competition looks like
The deeper story here is not that Anthropic is scared of science fiction. It is that frontier AI competition is entering a more self-referential phase. Models are no longer just products sold to customers. They are tools used to improve the labs that build the next round of models. Once that loop tightens, every gain in coding, experimentation, and research assistance has strategic weight.
That helps explain Anthropic’s mixed posture. It wants to move quickly enough to stay in the race, but not so recklessly that it loses the ability to argue for constraint. It wants to present itself as the serious adult on safety, while also protecting the advantages created by its own systems. None of that is especially shocking. It is what one should expect from a frontier lab facing both technical upside and strategic risk.
The key point is simpler. Anthropic is telling us that Claude is not just answering prompts and drafting snippets. It is already part of the machinery used to build what comes next. If that is true, then the successor problem has moved closer to the present. The worry is no longer whether AI might someday help create a stronger descendant. The worry is how far that process has already begun, and how candid the companies involved are willing to be about it.
A short close
Anthropic’s warning deserves attention, but not deference. The company may be right that frontier AI needs stronger brakes, tougher thresholds, and more than transparency theatre. It may also be defending its own lead. Those motives are not mutually exclusive. What matters now is that one of the firms closest to the metal is openly describing a world where AI is helping build the next AI. That is not a distant scenario any more. It is a governance problem arriving ahead of schedule.