Part of: this research — view all 2 rows
To resolve in the affirmative, a frontier AI developer must publicly report halting or pausing capability development (training or scaling) explicitly because its own published AI-research-automation threshold was reached. A declaration that a threshold has been crossed is insufficient without a corresponding halt, as are pauses attributed to other triggers like security incidents or misalignment. As of mid-November 2026, no qualifying halt has been reported. Recent incidents—such as OpenAI's July 2026 RL pause or Anthropic's pre-training delay—demonstrate that developers can and will interrupt expensive training runs, but these were security and alignment responses, not threshold-triggered pauses Pacing model development in an era of cyber-critical ....
The gap between reaching a threshold and halting training is best illustrated by Anthropic's recent disclosures. On November 10, 2026, Anthropic stated in a regulatory filing that it had determined it met its AI R&D threshold based on expected future progress Anthropic, PBC - S-1/A-2 Anthropic says it met its AI research threshold. However, this declaration was not accompanied by a halt in development. The company's leadership and public messaging have consistently emphasized "pacing, not pausing" Fable 5.2 Proves Pacing, Not Pausing (Nov 2026), focusing on implementing stronger safeguards, human-review requirements, and capability gating rather than a complete shutdown of training Anthropic says a key AI safety alarm has been triggered ... Pace the Frontier: Dario Amodei's Case for Slowing Down .... Anthropic has also noted that its willingness to pause depends heavily on other developers doing so in a verifiable manner When AI builds itself.
Over the next five years, it is highly likely that multiple developers will confront their automation thresholds. Evidence suggests that AI-driven research acceleration is already substantial: an independent assessment by METR estimated Anthropic's progress acceleration at roughly 2× (with roughly 80% attributed to AI assistance), and Anthropic's own internal capability index showed a 1.88× acceleration by early November 2026. While Google DeepMind recently found no significant end-to-end acceleration on its own metrics, its leadership expects the next twelve months to bring a steep inflection point because of recursive self-improvement. As capabilities grow, developers will repeatedly face consequential decisions about how to manage these acceleration thresholds.
Despite these rapid advancements, developers face intense competitive and commercial pressures that strongly discourage unilateral training halts. When thresholds are crossed, frameworks generally mandate the implementation of mitigations rather than an automatic cessation of work Anthropic's Responsible Scaling Policy Frontier Safety Framework 3.1 - Googleapis.com. Anthropic's Responsible Scaling Policy requires an affirmative safety case rather than a mandatory pause Responsible Scaling Policy, and Google's framework triggers an assessment and response plan Frontier Safety Framework 3.1 - Googleapis.com. While OpenAI's current Preparedness Framework explicitly commits to "halt further development" if its Critical self-improvement threshold is reached Preparedness Framework, the company is operating well below that rating today and has indicated it will replace the framework with a vaguer commitment to slowing or stopping systems it cannot sufficiently safeguard Pacing model development in an era of cyber-critical ... Pacing model development in an era of cyber-critical ....
While a strict halt is the exception rather than the rule, plausible pathways to a qualifying pause remain. A safety-conscious developer might temporarily delay a large training run specifically to finalize the governance, human-review protocols, or affirmative safety cases required by a newly crossed automation threshold, and explicitly attribute that delay to the threshold. Alternatively, OpenAI could reach its Critical threshold while its current strict language remains in force. Finally, mounting coordination efforts, such as the "Pacing the Frontier" initiative or future government-backed pacing regimes, could eventually provide the political and industry cover for a developer to pause and explicitly cite its own R&D threshold as the trigger. Weighing the anticipated timelines for AI R&D automation milestones against developers' structural incentives to deploy pacing mitigations rather than complete halts, I assign a 26% probability to a qualifying halt over the next five years.
Evaluating this question against anticipated timelines for AI R&D automation milestones and developers' structural incentives to deploy pacing mitigations rather than complete halts kept our estimate largely stable.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited