Anthropic says AI may soon double how fast it builds new AI, and its CEO says the industry should slow down. Here is what that could mean for tools you use.
Anthropic, the company behind the Claude AI system, says it may soon reach a point where AI speeds up its own development. In an amended filing dated November 2, made as it prepares to sell shares to the public, the company says it "may imminently meet" that threshold. It is still checking whether its models have already crossed it.
Anthropic's own safety policy sets the line. It is about two years of normal progress packed into one year. The company says Claude already writes and reviews code, runs experiments, prepares training data and tests the models that will replace it. Anthropic calls this "closing the loop."
The filing says the value of Anthropic's internal model use, measured at list prices, grew more than 250-fold from January to September 2026.
Anthropic co-founder Jack Clark wrote in mid-October that fully self-improving AI is now more likely to be possible in 2027 than in 2028. Rayan Krishnan, CEO of Vals AI, predicted it could arrive around July or August 2027. He also said chip design and other limits could hold back further speedups.
Not everyone sees a straight line. John Schulman of Thinking Machines said models still struggle to judge and check their own work. Writing more code, he said, does not make a team 100 times more productive. Google CEO Sundar Pichai said Google is "pretty far from the end state" of self-improving AI.
In a September essay, Anthropic CEO Dario Amodei wrote, "We must slow the pace at which we improve the capabilities of AI models." His reasoning is that safeguards need time to catch up. He proposed outside evaluators with ongoing access to top labs, and possible limits on how much companies use AI to improve AI.
Anthropic has committed to embedded evaluators. The wider limits need other companies and governments to agree. The filing admits the risk: these measures "could slow our capability gains relative to competitors that do not adopt similar constraints."
Anthropic has not said any current model has crossed its threshold. It expects people to stay in charge, and it is considering policy changes to keep them there. For everyday users, the likely effect is faster changes to AI tools and a louder debate over who sets the limits. How fast, and whether rivals follow, is not known yet.
AI tools may start improving faster than the rules around them, so expect quicker changes and a louder debate over limits.
World models, AI that learns physics from video, are being built into home robots and driverless cars, and your own videos may help train them.
Read more: AI that learns how the physical world works is coming to robots and carsA new Alibaba-led test of real coding requests found the best AI agent fully finished just 26.4 percent of tasks, so human review still matters.
Read more: AI coding helpers fully finish only about 1 in 4 real tasksA University of Chicago law professor asked Claude to fix page numbers. It deleted his syllabus and a year of notes instead. Dropbox saved him, not the AI.
Read more: An AI assistant deleted a law professor's year of lecture notes
