Part of: this research — view all 3 rows
Framing the Baseline The relevant comparison is the lag between competing frontier organizations, not the gap between closed and open-weights models. Because "a second organization" can be another closed lab, the historically tight race among top hyperscalers anchors the baseline. Epoch's estimated open-weights lag (about 4 months) or its cyber capability lag (6–7 months) The Epoch Brief - October 23, 2026 Open models lag state-of-the-art closed models by 4 months Open models lag state-of-the-art closed models by 4 months are less relevant than direct closed-lab competition.
Empirical Catch-Up Dynamics A historical reconstruction of integer-threshold crossings on Epoch's Capabilities Index (ECI) for 2025 and 2026 reveals a median catch-up lag of roughly 56 to 70 days https://epoch.ai/data/eci_scores.csv https://epoch.ai/data/eci_scores.csv https://epoch.ai/data/eci_scores.csv. Even today, the modest ECI gap between Anthropic's Claude Fable 5.2 and Google's Gemini 4 Carbon translates to roughly 80 days of progress at the current frontier improvement pace https://epoch.ai/data/eci_scores.csv. While Artificial Analysis's v5 index shows a widening open-weights gap https://artificialanalysis.ai/trends, this relies on rescaled scores and new evaluation methods, making temporal divergence hard to cleanly establish Announcing Artificial Analysis Intelligence Index v5. Consequently, the near-term baseline remains firmly in the two- to three-month range.
Compounding and Continual Learning Dwarkesh Patel predicts that when deployment feeds back into training, the returns to being ahead will accelerate and lags will widen 8 Predictions for the Era of Continual Learning. This specific continual-learning mechanism is not yet operating at the frontier; as of late 2026, no publicly known frontier model updates weights from live sessions. However, compounding advantages do not need to wait for live-deployment learning. Compute concentration among five hyperscalers is extreme, and AI-accelerated R&D is already yielding dividends. METR estimates approximately a 1.9-fold acceleration in Anthropic's capability improvement, largely attributed to internal AI use (reducing to ~1.4-fold when factoring in competitors) Independent Assessment of AI R&D Acceleration .... This suggests early feedback loops could begin stretching the leader's advantage meaningfully between 2027 and 2029.
Withholding and Safety Friction Structural friction is also pushing release lags longer. Frontier labs are increasingly holding back their best models for alignment testing and safety reviews before public release. METR notes that Anthropic is "holding onto more recent models for slightly longer before their public release" The Epoch Brief - November 11, 2026, and OpenAI has unreleased internal models that score significantly higher than released counterparts on internal benchmarks. While algorithmic efficiency improves by about 3x a year Algorithmic progress in language models and diffuses rapidly, safety pauses, export controls, and reported anti-distillation measures Disrupting a coordinated model-distillation campaign act as formidable counterweights to frictionless diffusion.
Final Assessment The median estimate of 78 days balances the robust empirical baseline of rapid closed-lab catch-up against the increasing pressures of internal holding and potential R&D automation. Anticipated increases in safety-driven release delays and the potential for compounding R&D advantages warrant a modest premium over the historical baseline, consistent with expectations for these mechanisms in related questions. The 10th and 25th percentiles are 28 and 48 days, respectively, allowing for continued rapid catch-up among closely matched frontier labs; the 75th and 90th percentiles are 135 and 240 days. The distribution is heavily right-skewed. The upper tail accommodates the plausible risk of a regime change by 2028–2029—where one lab achieves an R&D breakthrough that takes competitors six to twelve months to replicate—as well as definitional uncertainties, stricter statistical matching conventions Open models lag state-of-the-art closed models by 4 months, and the mechanical possibility of unmatched thresholds near the December 2029 deadline.
Set against related questions, this was adjusted slightly upward to maintain coherence with anticipated increases in safety-driven release delays and the potential for compounding R&D advantages.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited