Part of: this research — view all 3 rows
The Anchor and Patel's Mechanism Dwarkesh Patel predicts that the "era of continual learning" 8 Predictions for the Era of Continual Learning will force labs to deploy their smartest models earlier to avoid ceding deployment learning 8 Predictions for the Era of Continual Learning. Currently, this mechanism is not operating for frontier models. As of mid-November 2026, everything shipped at the frontier is retrieval or context injection with zero weight change. While non-frontier systems like Cursor Tab Improving Cursor Tab with online RL and Shopify Sidekick's continual learning loop (2026) update weights rapidly, frontier labs have not solved continual learning for their flagship systems. Consequently, the "deployment-learning prize" will likely exert little to no pressure on release gaps for most of the 2027–2028 window. Any compression below the established ~106-day anchor must instead come from ordinary competitive cadence.
Evaluating the Empirical Baselines The cleanest dated observation for an internal-to-external gap is Anthropic’s Mythos class, which saw about 106 days between internal use (2026-02-24) and general availability (2026-06-09). Recent datapoints suggest shorter intervals, though many are strictly lower bounds. OpenAI's GPT-6 Astra was in internal use at least 14 days before its release Covert underperformance on safety research tasks, though its actual gap was plausibly several weeks given earlier cyber determinations Pacing model development in an era of cyber-critical .... Google’s Gemini 4 Carbon reached its internal coding platform roughly 35–40 days before general availability Google Employees Praise Latest Gemini 4 Internal Model Google Tests Carbon, an Internal Gemini 4 Model for Coding. Conversely, Anthropic's Mythos 5.2 was accessible to external evaluators at least 47 days before release Claude Fable 5.2 & Claude Mythos 5.2 System Card, meaning the true internal gap was likely significantly longer.
Competing Pressures on Release Cadence While competitive pressure and leaderboard chasing push OpenAI and Google toward faster commercialization, safety and regulatory burdens exert a strong countervailing force. METR notes that Anthropic is holding onto its most capable models "slightly longer" before public release Independent Assessment of AI R&D Acceleration ... Independent Assessment of AI R&D Acceleration .... Unreleased models from both Anthropic and OpenAI already outscore their public counterparts (Fable 5.2 and Astra, respectively) on internal and external benchmarks Fable and Mythos 5.2: The System Card - by Zvi Mowshowitz. Furthermore, safety obligations are compounding: OpenAI has shelved models for failing alignment evaluations Periodic Summary of Internal Use Catastrophic-Risk ..., and Anthropic relies heavily on pilot internal usage to build alignment evidence before release Redacted Risk Report August 2026 - Claude. Regulatory friction, such as export-control suspensions and voluntary pre-release government access windows New AI Executive Order Calls for Frontier Model Security ..., will predictably extend timelines. Across related forecasting questions, an increasing consensus that safety and alignment friction will lengthen future release timelines lends modest additional weight to these pressures, rather than to competitive compression alone.
Conclusion and Distribution Shape Pooling across the three major labs, I expect the median gap for flagship models to land around 82 days. Google and OpenAI will likely pull the aggregate median down with typical gaps in the 40–80 day range, while Anthropic’s methodical alignment approach will keep its models in the 90–140 day range. The uncertainty distribution is heavily right-skewed, with a 10th percentile of 32 days, a 25th percentile of 52 days, a 75th percentile of 122 days, and a 90th percentile of 175 days. While true frontier deployment learning could theoretically compress gaps to a matter of weeks by late 2028, mandatory safety holds, dangerous-capability findings, or export restrictions could easily push individual model lags to many months. The sparse disclosure of internal-use dates What Should Frontier AI Developers Disclose About ... introduces further measurement uncertainty, likely biasing public resolutions toward longer, safety-relevant delays.
Set against related questions, the distribution was slightly raised to reflect an increasing consensus that safety and alignment friction will lengthen future release timelines.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited