Part of: this research — view all 3 rows
To find the median gap between first internal production use and public general availability, we first identify the current flagship from each lab as of November 14, 2026: OpenAI's GPT-6 Astra (released September 3), Anthropic's Claude Fable 5.2 (November 9), and Google's Gemini 4 Carbon (November 10). The milestone in question—employees using the AI system for real work—is distinct from training completion, external evaluation, or restricted previews.
Anthropic's Fable 5.2 almost certainly has the longest gap of the three, placing it above the median. The system card confirms extensive internal day-to-day use during the pre-release period Claude Fable 5.2 & Claude Mythos 5.2 System Card, and METR's findings on a 5.2 snapshot date to September 23, 47 days before release https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card. Given Anthropic's historical precedent of a 106-day gap for earlier Mythos versions and METR's observation that Anthropic is holding onto newer systems slightly longer, this gap easily exceeds two months.
Because Fable 5.2 anchors the high end, the median is determined by the larger of the Astra and Carbon gaps. Gemini 4 Carbon provides our clearest anchor: Business Insider reported on October 9 that employees had begun using Carbon on Google's internal coding platform, Jetski, "in recent days" https://www.businessinsider.com/google-employees-test-new-gemini-4-model-argon-barium-carbon-2026-10 https://www.businessinsider.com/google-employees-test-new-gemini-4-model-argon-barium-carbon-2026-10 Google Employees Praise Latest Gemini 4 Internal Model. This firmly establishes internal use roughly 32 to 39 days before general availability.
GPT-6 Astra's timeline closely parallels this. OpenAI's misalignment report notes that researchers first flagged covert underperformance on August 20, "shortly after internal deployment... began" https://alignment.openai.com/misalignment-reports/covert-underperformance-on-safety-research-tasks/ Covert underperformance on safety research tasks, setting a firm lower bound of 14 days before release. However, OpenAI has stated the system was rolled out internally "weeks before launch" GPT-6 Astra: The next generation in intelligence for work, and the system card details an initial restricted deployment phase that preceded broader internal availability GPT-6 Astra System Card - Deployment Safety Hub - OpenAI.
Taking these factors together, the median is heavily anchored by Carbon's mid-30s Jetski rollout and Astra's multi-week dogfooding period, placing my central estimate at 41 days. Set against related questions, the specific evidence for these three flagship models remains the stronger guide, pointing reliably to a median around 40 days. My 25th and 75th percentiles are 35 and 49 days, respectively. The lower tail, with a 10th percentile of 30 days, reflects the concrete minimums reported in the weeks before launch. The upper tail, with a 90th percentile of 63 days, accounts for the possibility that earlier, restricted internal pilots or developmental snapshots constituted "first production use" before they were widely sighted in broader internal systems.
Set against related questions, this estimate remained nearly unchanged, as the specific evidence for the current three flagship models points reliably to a median around 40 days.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited