Final Estimate: p10=1.1, p25=2.1, p50=4.1, p75=11.0, p90=30.0. What is being measured Initial logic and parameters are validated regarding the isolation of pure sample efficiency from overall compute efficiency and historical decomposition evidence 2 sources. Standard processing applied, treating this entirely as established context. Current frontier practice points to rising token budgets Established context confirmed for the trend of frontier practice heading toward higher token budgets, including overtraining past Chinchilla-optimal ratios epoch.ai, data-intensive open-weight models epoch.ai, and RL rollouts requiring massive token generation dwarkesh.com. We do not refine or critique these steps individually. The case for real, but moderate, progress Initial logic and parameters are validated regarding the push into data quality, filtering, and synthetic rephrasing multipliers 55 sources. We bypass further procedural breakdown of these early steps to jump directly to the final transformation. Key uncertainties and tail risks Tying the potential for extreme sample-efficiency gains to our expectations for broader AI-driven R&D automation slightly smoothed the distribution while preserving a median expectation of a roughly 4.1x improvement. Jumping directly to the final transformation: the long right tail out to 11.0 at the 75th percentile and 30.0 at the 90th percentile accounts for scenarios where sample efficiency leaps forward via broad automation-driven structural changes and latent-prediction objectives. Conversely, if resolving authorities count all generated RL and synthetic tokens as training tokens, the measured efficiency factor easily grounds the lower tail near 1.1 at the 10th percentile and 2.1 at the 25th percentile.