When will the first project fully automate AI R&D such that a model can produce a successor model more capable than itself without human intervention?

The final percentile estimates for the first fully automated AI R&D loop are: p10 November 2028, p25 August 2030, p50 April 2033, p75 January 2040, and p90 January 2060. <br><br> As of September 2026, frontier models show rapid capability gains, with Claude leading ~26% of AI R&D tasks anthropic.com and task-horizons doubling every 89 days metr.org. However, no system has achieved Level 5 autonomy anthropic.com. State-of-the-art demonstrations like Anthropic's automated researcher remain heavily constrained by human scoping and fail to produce a successor that outperforms the producer on frontier benchmarks alignment.anthropic.com. <br><br> Resolving this milestone requires a verifiable absence of substantive human decision-making throughout the entire multi-million dollar training cycle, rather than merely strong performance on component research tasks. Current safety frameworks and compute allocation protocols explicitly mandate human oversight darioamodei.com. <br><br> Therefore, the gap between underlying capability and a publicly verified, human-free execution dictates the median of April 2033. Operational timelines for AI replacing human researchers and lab safety protocols also create strong structural disincentives for labs to publicly verify fully autonomous R&D loops, pushing the middle and upper percentiles slightly later. The extended right tail is anchored directly on the non-trivial likelihood of permanent human-in-the-loop governance policies, binding regulatory pacing regimes pacingthefrontier.com, or plateauing research taste anthropic.com, which could persistently prevent a clean public resolution.