AI Is Beginning to Write Its Own Future, Anthropic Says
Anthropic says Claude now writes more than 80% of the code merged into its codebase, pointing to a future where AI helps build its own successors.
News
- Gemini Flaw Let Hackers Exploit Android Notifications
- AI Is Beginning to Write Its Own Future, Anthropic Says
- Data Centers Find Friendlier Ground in India, Survey Finds
- Supreme Court Proposes AI Rules for Courts, Bars Algorithm-Only Judicial Decisions
- SAP Labs India Selects 16 Startups for 2026 Startup Studio Cohort
- Nearly Half of Workers Now Spend More Time Managing AI Than Doing Tasks, Study Finds
Artificial intelligence is taking on a larger role in building new AI systems, according to Anthropic, which says Claude now writes more than 80% of the code merged into the company’s codebase.
In a report published by The Anthropic Institute, the company said AI tools are accelerating software engineering and research work inside Anthropic, narrowing the gap between AI assistance and a future in which AI systems could help design their own successors.
“For most of AI’s history, humans drove every step in its development cycle. But at Anthropic, we are delegating a growing share of AI development to AI systems themselves, which is speeding up our work,” the report said.
Anthropic described the trend as a possible step toward “recursive self-improvement,” a scenario in which AI systems help build more capable future versions of themselves.
The company said that has not yet happened, but warned that institutions may have less time to prepare than they assume.
The report combines public benchmark data with internal metrics from Anthropic’s engineering and research teams.
As of May 2026, more than 80% of the code merged into Anthropic’s codebase was authored by Claude. Before Claude Code launched in research preview in February 2025, the figure was in the low single digits.
Anthropic said the typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024. It cautioned that lines of code are an imperfect productivity measure and likely overstate the true gain, but said the data still points to a significant acceleration in software development.
In a March 2026 poll of 130 employees across Anthropic research teams, the median respondent estimated producing about four times as much output with access to Mythos Preview as they would have without AI assistance.
Anthropic said the true uplift was probably lower, but broadly consistent with other signs of faster technical work.
The company cited examples of AI completing work that may otherwise have remained unfinished. In April 2026, Claude shipped more than 800 fixes that reduced one class of API errors by a factor of 1,000.
The engineer overseeing the work estimated that a human would have taken four years to complete it.
“I started leaning hard into Claudifying about a year ago,” one Anthropic employee said in comments included in the report. “That’s been a crazy adventure and it’s now been ~5 months since I last wrote any code myself.”
Anthropic said Claude is also becoming more capable of handling open-ended engineering tasks. Its success rate on such tasks reached 76% in May, up 50 percentage points in six months.
In one example, Claude identified the cause of a training-job failure in about two hours, work Anthropic said would normally take engineers two to three days.
The company also pointed to progress in AI-assisted research. It said Claude Mythos Preview achieved about a 52-fold speedup in a defined research-coding task in April, compared with about threefold for Claude Opus 4 in May 2025.
A skilled human researcher would typically need four to eight hours to reach a fourfold improvement, Anthropic said.
Despite the gains, Anthropic said humans still retain an advantage in deciding which problems matter and in judging broader research direction.
“The comparative advantage of humans as of right now is still in seeing the bigger picture and thinking beyond the confines of the immediate task,” the report said.
Anthropic said human roles may increasingly shift from writing code and running experiments to setting goals, reviewing outputs and overseeing AI systems.
It outlined several possible futures. Progress could slow because of technical or infrastructure constraints. AI could continue to automate larger parts of research and engineering while humans retain strategic control. Or, in the most uncertain scenario, AI systems could become capable of designing successor systems with limited human involvement.
Anthropic is calling for frontier AI developers to establish a coordinated and verifiable way to slow or pause development if advanced systems begin improving themselves faster than society can manage the risks.
The company also warned that unilateral pauses may have limited effect if less cautious rivals keep advancing.
“If it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that would likely be a good thing,” Anthropic said.
Anthropic said it plans to convene discussions with policymakers, researchers, civil society groups and other AI developers on recursive self-improvement, coordination and oversight.
