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)?

This question resolves only when a frontier developer formally publishes an AI R&D automation index rating at least one measured subset of work at Epoch's AL5 ("fully autonomous") or an equivalent top level. As of October 2026, Anthropic is the only developer publishing such an index. On its August 2026 data, AL5 stood at zero anthropic.comanthropic.com. While Claude "leads" (AL4) 26% of Anthropic's AI R&D work and at least "collaborates" (AL3) on more than 90%, it operates fully autonomously in none anthropic.com. Other labs use different metrics—OpenAI sorts agent tokens by Epoch's six-phase taxonomy without autonomy levels and stresses that humans still set priorities openai.comopenai.com—or publish no index at all.

The jump from AL4 to AL5 represents a deliberate governance choice rather than just a capability threshold anthropic.com. In Anthropic's AL4 framework, the AI can monitor, design, test, and write fixes, but an engineer still reviews the change and decides whether to ship. Reaching AL5 requires trusting the model to deploy to production without human oversight. Following recent agent security incidents—such as the compromise of OpenAI's research infrastructure—and amid public calls for pacing the frontier, labs face strong reputational and security disincentives to removing the human from the loop. The presence of embedded METR evaluators and commitments under the White House Accord further elevate the political and safety costs of declaring any subset of work fully autonomous.

Despite these frictions, there are plausible paths to an AL5 rating emerging by late 2027 or 2028. Anthropic's task basket is highly granular, spanning 378 leaf categories anthropic.com. Moving a single narrow, low-risk category—such as pipeline monitoring, eval runs, or flaky test fixes—to auto-deploy would be enough to resolve the question. Furthermore, Anthropic's ratings rely on a Claude judge model that exhibits only 59% exact agreement with humans, meaning judge noise or borderline categorizations could result in an early AL5 leaf as AL4 tasks proliferate anthropic.com. Other frontier labs, acting on transparency commitments, might also adopt the AL scale and report autonomous subtasks of their own.

The long right tail, extending into the 2030s, reflects deep structural uncertainties about measurement and organizational policy. There is no guaranteed publication schedule for these indexes; Anthropic has only committed to periodically "rebuild the basket" and "re-version" the numbers anthropic.comresultsense.com. Crucially, labs could permanently enforce human sign-off for all R&D tasks, functionally capping the automation scale at AL4 indefinitely. Measurement series in this space are also fragile and frequently discontinued—as demonstrated by the abandonment of METR's time-horizon suite—introducing the risk that the required AL5 reporting metric is simply retired before it can ever be achieved.