probability: 30%

rationale:

Status Quo and the Benchmark Gap The final estimate is 30%. As of August 2026, no Google DeepMind (GDM) report or model card has declared an ML R&D alert threshold crossed. The gap between current capabilities and the threshold remains stark. GDM's Frontier Safety Framework (FSF) v3.1 defines a "rule-out" threshold for the Critical Capability Level (CCL) at 90% pass@1 on its internal research-engineering benchmark (GRB), with the alert threshold set "marginally earlier" storage.googleapis.com. The August 2026 Gemini 3.7 Flash FSF report puts the model at 27% on this evaluation—and even a maximally generous adjustment for bugged tasks would only reach 47% storage.googleapis.com. GDM explicitly notes that the model lacks the independence to chain tasks into end-to-end research workflows deepmind.google. Consequently, across three consecutive major Gemini reports, the measured capability has remained "substantially" or "comfortably" below the alert threshold 3 sources.

Release Cadence Constraints Reaching the alert threshold by the end of 2027 requires aggressively closing this quantitative gap within roughly 16 months. Google confirmed Gemini 4 entered pre-training in July 2026, implying expected general availability around mid-2027 futuresearch.ai. This release cadence suggests only one or two genuinely new frontier-generation model cards (plus interim updates) will land before the deadline. While GRB pass rates grew from roughly 16% (Gemini 3.1 Pro) to 27% (Gemini 3.7 Flash) in six months storage.googleapis.com, reaching the ~80–90% range by late 2027 requires several more such doublings into the hardest tail of a long-horizon benchmark.

External Evidence and Peer Labs Evidence from peer labs and independent evaluators reinforces this timeline bottleneck. Anthropic's August 2026 risk report concludes its models do not cross its own automated AI R&D threshold—which is significantly weaker than Google’s—and notes that despite productivity uplifts, internal research velocity has not yet doubled anthropic.com. Similarly, METR's evaluations find that while agents can write complex PRs and optimize metrics, they still lack high-level research judgment, and companies have not reported dramatic, lab-wide speed-ups in overall progress attributed to automation metr.org. However, the rapid progress of narrow AI research tools at these peer labs indicates that piecemeal acceleration is actively unfolding.

Alternative Pathways to a Crossing Despite the formidable capability gap, several factors keep the probability out of the extreme tail. Most importantly, GDM has a demonstrated willingness to disclose alert crossings when they occur, having already reported reaching them for cyber and CBRN uplift deepmind.google. Furthermore, the question asks only about the alert threshold, not the full CCL. The ML R&D acceleration CCL is inherently observational (hinging on whether AI "has been used to accelerate AI development... substantially from historical rates") deepmind.google. This means an internal-deployment judgment call regarding Google's own R&D speed-up could trigger an alert without requiring pure benchmark saturation storage.googleapis.com. Assessing the potential for these softer observational triggers alongside the rapid progress of narrow AI research tools at peer labs slightly raises our estimate. Ultimately, weighing the multiple opportunities for a new model release and the softer, judgment-based "acceleration" trigger against the very wide measured capability gap and tight model-release timeline, the likelihood of a threshold-crossing statement before 2028 is estimated at 30%.