Question
When will any frontier developer's own published AI R&D automation index rate any measured subset of its AI R&D work at the fully autonomous level (Epoch AI's Automation Level AL5, or the top level of a successor scale)?
Resolution Constraints and Current Status The resolution criteria are strict: a frontier developer must publish its own automation index or measurement series rating a measured subset of its AI R&D work at AL5 (fully autonomous) or the top level of an equivalent scale. Product descriptions calling a system "autonomous," system-card capability claims, METR statements, and Google DeepMind's delegated-pull-request metrics do not count. As of mid-November 2026, Anthropic's R&D Automation Index (published 2026-09-17) is the only instrument using the AL0–AL5 scale. It found that Claude "leads" (AL4) 26% of R&D work but that no measured subset is fully autonomous anthropic.com. Anthropic plans to "rebuild the basket periodically and re-version" its numbers but specifies no cadence anthropic.com, and no second edition has yet appeared anthropic.com.
Capability Drivers and the Low Target Bar The primary mechanism pushing toward an early resolution is the granularity of the measurement. The current index is a tree containing 378 leaf task categories anthropic.com; only a single leaf needs to reach AL5 to trigger resolution. Narrow operational workflows, such as incident postmortems or infrastructure defect diagnosis, are highly plausible early candidates, well before core research direction-setting achieves autonomy. Furthermore, capabilities and R&D automation are rapidly accelerating. The AL4 share jumped from under 1% in February to 26% in August 2026. METR's recent independent assessment found roughly 2× acceleration with no bottleneck expected in the next 12 months, and Anthropic's internal agent hours have skyrocketed. Anthropic has also formally determined that it met its RSP AI R&D threshold based on expected progress, promising "further reports" that could serve as a vehicle for an index update.
Governance Frictions and the Definition of Autonomy Conversely, AL5 sets a demanding standard for complete autonomy. In Anthropic's worked example, the AI must monitor for failures itself, scope the problem, design and implement a fix, test it, and deploy it to production—all without human involvement anthropic.com. Human initiation and shipping approval firmly separate AL4 from AL5. Furthermore, resolution is as much a question of internal governance and public reporting as it is of raw capability. Anthropic is considering RSP changes that would formally require human review and approval for certain R&D decisions 2 sources. A deliberate "human-in-the-loop" policy across task boundaries could persistently cap every leaf at AL4, even if underlying capabilities could theoretically support AL5.
Timeline Dynamics These cross-pressures dictate the percentile spread. An early resolution (by Q1 2027) is realistic if Anthropic's next report or promised update catches a narrow operations leaf crossing the threshold, perhaps assisted by variance in the Claude judge model used to rate tasks implicator.ai. The median falls in late 2027, allowing for a few internal measurement cycles and the continued compounding of capabilities. The long right tail, extending to mid-2030, accounts for the substantial risk that public AL-style reporting is deprecated, or that strict human-in-the-loop oversight mandates keep every measured subset formally short of the fully autonomous tier.
Set against related questions, the forecast remains unchanged, reflecting consistent timelines where granular subset automation precedes broader lab-wide capability declarations .
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited