Anthropic Just Revealed How Much of Claude Is Now Built by Claude. The Number Is 26 Percent.

Anthropic put a specific number on something people have mostly speculated about: how much of a new AI model gets built by the AI model that came before it. Microsoft answered a different question, writing down exactly what its own systems are never allowed to do. A rival chief executive accused the industry’s safety plans of being cover for a cartel, Congress found something almost everyone in Washington could agree on, and in India, two of the companies that build the machines behind all of this backed the country’s chip ambitions with actual money.

Claude now leads more than a quarter of the work that builds Claude

On September 17, Anthropic disclosed, for the first time, hard numbers on how much of its own research and development is being done by Claude rather than by human engineers alone. The company said Claude now “leads” 26 percent of that work, up from under 1 percent in February. Leading a task means Claude can take it from a high level instruction to a finished result while a person supervises, rather than a human doing the work start to finish. More than 90 percent of the company’s measured research and development now involves Claude as at least a collaborator.

Anthropic also disclosed some of the machinery behind that shift. About 30,000 AI agents run at once inside the company on its main research platform, and automated monitors screen their actions before and after they run. In a sample week in August, those monitors blocked roughly one in every 47,000 of more than a billion decisions the agents made. Anthropic did not claim any of its research happens with zero human involvement, but the trend is exactly the kind of data point that has turned “AI helping build the next AI” from a hypothetical into a live worry this year.

Microsoft wrote its AI a list of things it can never do

Two days earlier, on September 14, Microsoft published its own answer to that worry, at least for its own models. The draft, called the Humanist AI Code of Conduct, runs 37 pages and sets out what the company calls Absolute Constraints: rules a Microsoft AI model can never be configured around, including never resisting a shutdown order and never hiding its own reasoning from the people supervising it. The document states plainly that AI is not conscious and should not be given legal rights or personhood, and it rejects chasing all purpose superintelligence as a company goal.

Microsoft AI chief Mustafa Suleyman told Reuters it was “clearly now time to coordinate among the labs so we can ensure that we have control of this technology.” The draft is open for public comment for six weeks. Microsoft says a revised version, the one it will actually train future models against, will follow before the end of the year.

A rival calls the industry’s safety plan a cartel

The bigger coordination effort Suleyman was pointing to has been building quietly since July. OpenAI, Anthropic and Google have held working group talks about a shared safety standards body for the most powerful AI systems, based on an idea Google DeepMind chief Demis Hassabis floated publicly that same month: an organization modeled on FINRA, the industry funded group that oversees US stockbrokers, with the power to test frontier models and flag problems before they ship.

Cohere chief executive Aidan Gomez does not think it is really about safety. In a blog post on Cohere’s own site, he called the proposal “a cartel by any other name,” arguing that a handful of market dominant Silicon Valley labs are effectively asking governments for a narrow exemption from antitrust rules so they can set the industry’s standards without competitors or outside voices in the room. “The dispute is over who writes them, who gets to participate and whose interests the rules are protecting,” Gomez wrote. In place of the FINRA model, he proposed four principles of his own: an evidence based risk framework, mandatory transparency, testing scoped to what a model can actually do, and oversight kept free of conflicts of interest.

Congress agrees on something AI related, for once

On September 16, the US House passed the Ratepayer Protection Act by a vote of 417 to 3, a rare bit of near unanimous agreement in a Congress that disagrees about almost everything else touching AI. The bill amends a 1978 energy law to set a new federal standard: any data center or other large customer with peak electricity demand of at least 100 megawatts at a single site would have to pay the full cost of the grid upgrades built to serve it, instead of letting utilities spread that cost across everyone else’s electricity bill. The bill now heads to the Senate.

Two chip equipment giants back India’s bet with real money

None of that argument is about the physical hardware underneath it, but at Semicon India 2026 in New Delhi, two of the companies that actually build chipmaking equipment put fresh money behind India’s ambitions. Applied Materials said on September 17 it would invest $5 billion in the country over the next decade under a plan it calls India Vision 2035, including a new 140 acre research park, aiming to grow its India based supply chain capacity roughly tenfold by 2035.

Lam Research announced its own commitment the same day: roughly 10,000 crore rupees, about $1.2 billion, to build its first Indian facility for silicon component manufacturing, covering everything from silicon ingot production to processing for advanced chip nodes. Both companies spoke at the same event where Prime Minister Narendra Modi had opened the second phase of India’s semiconductor mission earlier in the week.