What Has to Happen Resolution requires OpenAI, Anthropic, or Google DeepMind to state publicly that a model of its own has crossed, or is expected within the next generation to cross, its automated-AI-R&D capability threshold. Crucially, hedged language such as "plausible," "may," or "could" is insufficient—an explicit crossing determination or a firm forward-looking expectation is required. As of late August 2026, none of the labs have reached this mark. Notably, OpenAI's recent pause of Astra in August 2026 was a cybersecurity determination, explicitly not an AI Self-improvement one openai.com.

Anthropic (The Nearest Mover) Anthropic appears the most likely candidate for an early disclosure. Its Responsible Scaling Policy v3.4 sets the bar at either full substitution for its research staff or a sustained doubling of the rate of AI progress anthropic.com. While its August 2026 Risk Report rated the threat model as "Low" and stated current models do not meet the criteria, the report flagged saturated evaluations and early signs of acceleration, warning that models "may cross this threshold in the coming year" anthropic.com. Because Anthropic issues semi-annual risk reports, its next disclosure window around February 2027 serves as a natural locus for the earliest probabilities, capturing the 9% chance of an early Anthropic crossing and anchoring the 10th percentile at March 1, 2027.

OpenAI and Google DeepMind OpenAI's Preparedness Framework v2 has a seemingly lower hurdle, defining "High" risk in AI Self-improvement as the equivalent of providing every researcher a mid-career engineering assistant. Yet the recent GPT-5.6 family (Sol, Terra, Luna) failed to reach this mark deploymentsafety.openai.com. Furthermore, OpenAI's recent announcement that it is replacing this framework altogether injects substantial uncertainty, potentially shifting the threshold or delaying public disclosures openai.com. Google DeepMind remains the furthest from its respective bar; the recent Gemini 3.7 Flash model scored only 27% on internal capability benchmarks against a 90% rule-out threshold and still lacked the independence required for end-to-end research workflows storage.googleapis.com.

Institutional Hurdles and Shifting Definitions While rapid model progress pushes toward near-term resolution, severe institutional disincentives pull in the opposite direction. Declaring that a model has crossed these thresholds triggers immense operational burdens, including mandatory development pauses, extreme security upgrades, and rigorous board approvals. Consequently, developers have strong incentives to continually refine or rewrite their frameworks to avoid prematurely triggering these mandatory R&D pauses. This dynamic is already visible in practice, with labs modifying threshold definitions precisely as model capabilities approach 2 sources. Factoring in the probabilities of related early-alert safety triggers alongside our broader estimate that labs will likely adjust safety frameworks to avoid premature threshold crossings establishes slightly wider tails and a later timeline for the overall distribution. Additionally, persistent structural bottlenecks in models exercising autonomous research judgment remain a tangible obstacle technologyreview.com.

Final Assessment Balancing the steady cadence of model releases and escalating capability warnings against institutional reluctance and strict measurement standards places the median estimate at July 1, 2029, with a 25th percentile of January 1, 2028. An early resolution relies on Anthropic or OpenAI hitting early-alert safety triggers and upgrading their currently hedged rhetoric to an explicit "expected next generation" statement in 2027. Conversely, the long right tail—extending to January 1, 2032 for the 75th percentile and January 1, 2036 for the 90th percentile—accounts for the high likelihood of persistent "moving the goalposts" scenarios where developers modify frameworks to avoid mandatory R&D pauses, prolonged capability plateaus, or stubborn structural limits in automating AI research and development.