The target is the latent true median gap from the first internal production use by employees to general availability for flagship models from Anthropic, OpenAI, and Google in 2027–2028. The most robust historical anchor is Anthropic's Mythos sequence, which established a gap of roughly 106 days 44 sources. We estimate the 2026 median for this gap at around 68 days . However, accelerating competitive cadence and shrinking training-to-release lags must be weighed against intensifying governance.

Internal and third-party evaluations, periodic risk reports anthropic.com, EU AI Act enforcement digital-strategy.ec.europa.eu, and upcoming state-level audit mandates ilga.gov require models to clear heavy pre-deployment gates 2 sources. These incoming regulatory and audit requirements will structurally lengthen pre-release testing. Google also has a persistent pattern of long preview-to-GA staging mashable.com. Blending faster competitive cadences with these slower regulatory rollouts centers the expected median around 90 days, with an interquartile range from 65 days (P25) to 125 days (P75). The theoretical incentive for a race to the bottom via continual learning dwarkesh.com is not yet operating at the frontier, as weight-updating loops remain restricted to smaller models 2 sources. The left tail (P10 of 45 days) accounts for labs bypassing early-snapshot testing to deploy rapidly, while the right tail (P90 of 165 days) reflects the risks of substantial export-control friction or complex government bottlenecks intentionally keeping models internal for extended R&D.