Anthropic Says Claude Now Does 26% of the Work Building Its Next Model — and the US Just Cleared Its Next Release

Anthropic disclosed this week that its current models are now doing roughly 26% of the engineering work that goes into building its next-generation models — up from effectively zero as recently as February. The disclosure landed the same week the US Department of Commerce lifted restrictions on Anthropic's ability to release new AI models, giving the company a clearer runway just as the pace of "AI building AI" crosses from anecdote into something the company is now willing to put a number on.

The number that matters

Associated Press reported Thursday that Anthropic's models are now contributing meaningfully to the research and engineering pipeline that produces their successors. The 26% figure is not a claim that Claude is autonomously designing and training the next Claude — Anthropic was careful to say the model is not yet working completely independently — but it is a measurable share of the work going into the next generation, and it was effectively zero in February.

That delta is the story. Six months ago, the honest answer to "is AI building AI" was "not in any meaningful way." Today, Anthropic is telling the AP that more than a quarter of the work on its next model comes from its current models. That is not autonomy; it is a trend line.

What the 26% actually means

Anthropic framed the disclosure as part of a push for better public measurement of how much frontier-model development is being done by AI. The company said roughly 30,000 agents were performing research and engineering work as of August, and it has called for standardized industry metrics showing how much of frontier-model development is AI-driven.

The distinction matters because claims that AI is "building itself" can easily outrun what the evidence actually shows. Anthropic is arguing for better measurement precisely because the numbers are moving and the language around them is sloppy. This is the company that wrote, in the same disclosure, that better public measurement is needed as systems move closer to forms of recursive improvement that could become harder for humans to supervise.

The regulatory backdrop: Commerce clears the next release

The disclosure landed alongside reporting that the US Department of Commerce has lifted restrictions on Anthropic's ability to release new AI models — a regulatory green light that removes one source of uncertainty ahead of whatever comes after Claude Fable 5.1.

For a company that has publicly committed to responsible scaling and has been forthright about the risks of its own models being weaponized, the regulatory clearance is not the headline — but it is the context. Anthropic can now ship without the overhang of export-control scrutiny on the next release.

Why this is the Anthropic recursive self-improvement story for September 2026

This blog has been tracking the recursive self-improvement question across multiple posts this month. On September 18, AIPress published a piece arguing that recursive self-improvement had crossed from thought experiment to something partially observable. Anthropic's 26% disclosure is the most concrete data point yet in that argument.

It is not proof that a self-improving loop is running. It is proof that the ingredient that matters most — AI systems doing real engineering work on their successors — is now present at a measurable scale, and that the company building the models is itself calling for better public metrics to track exactly this.

The honest caveat

These are vendor-chosen figures, not independent benchmarks. Anthropic defines what "26% of the work" means, and it has an incentive to frame the trend as manageable rather than alarming. The number is directional, not definitive.

That said, the direction is what matters for anyone tracking the AGI timeline. A quarter of next-generation model work being done by current models, up from zero in six months, is a trend that does not require anyone to believe in a hard takeoff to take seriously.

What to watch next

• Whether Anthropic publishes the standardized metrics it is calling for — without them, the 26% figure is a single data point, not a series.
• Whether the 26% grows in the next disclosure, and whether Anthropic starts attributing specific engineering tasks (code review, architecture proposals, test generation) rather than an aggregate number.
• Whether other frontier labs start publishing comparable figures. Right now, Anthropic is the only one putting a number on it.

Sources: Associated Press (September 18, 2026); Anthropic disclosure reported by AP; US Department of Commerce export-control clearance reported by ABC News (September 19, 2026); AIPress recursive self-improvement coverage (September 18, 2026).