AI Is Accelerating AI: Anthropic’s Pause Proposal and Three Future Scenarios

AI Is Accelerating AI

Anthropic’s long‑form article "When AI builds itself" combines public benchmarks and never‑before‑released internal data to demonstrate that AI systems are increasingly capable of building the next generation of AI, and that this acceleration is occurring faster than most expectations.

Rapid Progress Metrics

The duration of tasks that AI models can complete independently is roughly halving every four months (previously every seven months). In March 2024 Claude Opus 3 finished a software task that normally takes a human four minutes; a year later Claude Sonnet 3.7 handled a 1.5‑hour task, and by the following year Claude Opus 4.6 managed a 12‑hour task. Extrapolating, tasks that currently require days of skilled engineering could fall within AI’s reach within the year, and by 2027 AI may handle work that presently takes weeks.

Public benchmarks echo this trend: SWE‑bench scores rose from single‑digit percentages to near‑saturation within two years, and CORE‑Bench success climbed from ~20% in 2024 to full coverage fifteen months later. METR measurements show Claude Mythos Preview sustaining at least sixteen hours of continuous work, hitting the upper limit of the test suite.

Anthropic’s Dual Workflow

Anthropic splits model development into engineering (code, infrastructure, training supervision) and research (experiment design, result interpretation, direction setting). In both realms the picture aligns: Claude can take a vaguely described problem, devise a solution, and execute it with only a goal from the human, while for well‑defined experiments Claude matches or exceeds skilled human performance. The remaining gap lies in judgment—deciding which problems merit effort.

Code Generation Dominance

By May 2026 over 80% of code merged into Anthropic’s repositories originated from Claude, up from single‑digit percentages before Claude Code’s preview in February 2025. Engineer output surged: a typical engineer’s daily merged lines in Q2 2026 were eight times the 2024 level, with most code authored by Claude and engineers shifting to guidance and review. An internal survey of 130 researchers reported a median four‑fold productivity boost when using Mythos Preview.

Claude’s code quality has reached parity with human engineers, and an automated Claude reviewer now screens changes for bugs, security issues, and defects, intercepting roughly one‑third of bugs that would otherwise reach production.

Improved Decision‑Making

Claude’s success rate on open‑ended tasks rose to 76% in May 2026, a 50‑point increase in six months. In a study of 129 human judgment moments, the best Claude model gave better choices than humans 51% of the time in November 2025, rising to 64% by April 2026.

Three Possible Futures

1. Trend stalls while AI capabilities are already widespread; growth may follow an S‑curve, and breakthroughs could require new architectures or supply‑chain constraints.

2. Continued compound efficiency gains —AI‑driven automation multiplies organizational output, enabling a 100‑person team to perform work comparable to tens of thousands, though bottlenecks shift according to Amdahl’s Law.

3. Recursive self‑improvement —AI systems design and refine their successors, making development speed dependent solely on compute. Human roles shrink to supervision and verification, and the technology could spill over into other scientific domains.

Pause Option

Anthropic argues that a verifiable slowdown or temporary pause could buy society time to address AI’s profound impact, but without global coordination such measures may merely let the least cautious actors catch up, increasing overall risk. Implementing a trustworthy pause would require multi‑nation labs to agree on verifiable stop conditions, akin to arms‑control verification mechanisms, yet the urgency of AI development leaves little time for such infrastructure.

Call to Dialogue

Anthropic plans to convene policymakers, researchers, civil society, and other AI firms in the coming months to discuss the outlined questions—especially recursive self‑improvement and coordination mechanisms—and will publish the outcomes publicly, inviting broader participation.