Anthropic published internal measurements in a blog post asserting that its model Claude now authors or co‑authors about 80 percent of the company’s code as of May 2026. That is a sharp rise from under 5 percent before the preview release of Claude Code, and the company said the trend could indicate systems that “design and refine themselves.” The report thrust the theoretical idea of recursive self‑improvement (RSI) back into public discussion, further splitting the AI community between those urging drastic precautions and those warning that exaggerated fears could harm beneficial innovation.
Measurements and concrete results
Anthropic measured AI‑attributable gains in software‑development productivity and extrapolated several future scenarios. According to the blog:
- Claude’s authorship or co‑authorship accounted for about 80 percent of the company’s code by May 2026. (In April, OpenAI president Greg Brockman said OpenAI models were authoring or co‑authoring a similar share of that company’s code.)
- After the Claude Mythos Preview launch, in the second quarter of 2026 each engineer contributed eight times more lines of code per quarter than in the first quarter of 2023, when Claude first launched.
- In April 2026 the company shipped more than 800 API fixes, reducing API errors roughly 1,000‑fold; engineers estimated these fixes would have taken humans working alone about four years to complete.
Anthropic also reported steady improvement in AI‑written code quality. The company used a large language model to classify code issues as trivial, routine, substantial, or open‑ended, and recorded performance gains from September 2025 to May 2026:
- trivial problems: less than 80% → about 90%
- routine tasks: 65% → 90%
- substantial tasks: under 40% → over 80%
- open‑ended problems: less than 20% → 76%
Three future scenarios
Based on these observations, Anthropic outlined three possible futures for AI and software engineering:
- AI remains less capable than the best human engineers.
- Anthropic’s preferred scenario: AI‑assisted software engineering continues to accelerate, but humans retain control over model research and development.
- AI reaches the point of being able to improve itself, i.e., recursive self‑improvement.
Reactions and points of contention in the AI community
The RSI idea is not new, but Anthropic’s report raised its profile and prompted a range of responses. OpenAI commented that it also sees early signs of RSI in today’s systems, and the Japanese research group Sakana AI launched an RSI Lab dedicated to studying self‑improving AI.
Many observers stressed the gulf between agentic coding—where agents respond to human engineers who organize, direct and evaluate efforts—and genuine ongoing self‑improvement, where systems handle the entire research and development process. Arun Rao, an adjunct professor at UCLA, said he expects a longer path than Anthropic anticipates. AI policy researcher Miles Brundage described himself as relatively skeptical about RSI. Matthew Barnett, co‑founder of MechanizeWork, pointed to data and compute bottlenecks as obstacles.
Others criticized the marketing tone of Anthropic’s framing. Wharton professor Ethan Mollick noted there is “some navel‑gazing, some marketing, and a lot of very sincere beliefs” in Anthropic’s presentation. Tech analyst Michael Spencer observed that the recent large seed funding rounds have often gone to AI startups focused on this trend.
Historical context
The notion of recursive self‑improvement goes back to early ideas of an “intelligence explosion,” most famously articulated by I. J. Good in 1965, who argued that a sufficiently advanced machine intelligence could improve its own design and rapidly surpass human intelligence. In the 2000s and 2010s Eliezer Yudkowsky (Machine Intelligence Research Institute, UC Berkeley) formalized RSI as a central concern of AI alignment research. The topic reentered mainstream research with the rise of large language models and AI‑assisted coding; research groups including the Chinese Information Processing Laboratory have proposed benchmarks such as the Meta‑Agent Challenge to evaluate RSI capabilities.
Why it matters
Like the prospect of artificial general intelligence (AGI), RSI currently appears distant: substantial work and likely breakthroughs are required for present systems—those that multiply human productivity in software development and other fields—to evolve into systems that oversee, design and engineer their own improvements in a recursive loop. At the same time, today’s systems already have concrete, near‑term impacts on productivity.
The AI community is divided over how talk of catastrophic futures should influence policy and public perception. Science‑fiction scenarios can be effective in attracting attention or funding, but realistic assessments of possible futures are necessary to make practical progress and to craft appropriate regulation.
Comment on the research‑pause suggestion
Anthropic’s blog revived discussion of a global, temporary pause in AI research. While the company did not call for a complete halt to all research, the idea put the notion of a pause back on the table. Critics of such pauses argue that they empower alarmism and could be counterproductive; proponents of regulation insist that dangerous applications be governed while fundamental research continues as appropriate.
In sum, Anthropic’s data have reignited debate: they document substantial gains in AI‑assisted software productivity, but whether those gains will culminate in true recursive self‑improvement remains an open technical and policy question that the field continues to grapple with.
