Anthropic says Claude already helps build its own successor, and a 2027 forecast looms

What Anthropic has admitted

As Phoenix New Media Tech (ifeng Tech) reported, citing the Wall Street Journal, Anthropic has publicly acknowledged that its Claude model is deeply involved in building its own successor. Researchers call this recursive self-improvement: an AI improves itself, the stronger AI improves the next one, and the loop repeats.

In theory, once that loop closes, progress could move faster than anything people have experienced. Some see a technology explosion in that. Others worry humans could lose control of the pace without noticing.

What the company's own report shows

In a report titled When AI Builds Itself, Anthropic said that most of the code merged into its production codebase is now written by Claude. When Claude Code launched, that share was only a small fraction. Engineers merge far more code per person than they did a couple of years ago.

Some engineers have not hand-written code for months. Their job is to manage and review the output of several Claude copies at once. Human review has become the bottleneck, because the AI writes faster than people can check.

The company also describes a growing share of its research and development as AI-led: humans set a rough direction, the AI does most of the work, and people supervise. Anthropic stresses that no research area is fully autonomous yet.

Why the date moved to 2027

Co-founder Jack Clark had earlier given better-than-even odds that AI could develop a successor model without human help by the end of a later year. He has now moved that forecast to 2027. He stressed that it is his own judgment, not a company product goal.

One input was a new benchmark called TasteVal, which tests whether an AI can pick promising experiments and read their results, a kind of research "taste." Recent models seem to be improving on it much faster than earlier ones. But none of the AI-submitted research proposals counted as a true invention. The AI was better at tuning and combining ideas than at starting new ones, and the researchers warned that the test is small and the human comparison group was not made up of top scientists.

Brakes, doubts and what to take from it

Anthropic's report argues the world may need a verifiable global way to slow or pause frontier AI, since one lab stopping alone would just hand the lead to others. It compares this to nuclear arms control but admits that training a model is far easier to hide than a missile silo. A public letter backed by many AI practitioners, including senior staff at Anthropic and OpenAI, called for government coordination on slowing down.

Skeptics, including NYU professor emeritus Gary Marcus, say the numbers show faster programming, not true self-evolution. And because much of the data comes from Anthropic itself, outsiders can't easily check it.