An AI model used to live 43 months before being switched off — models launched in 2025 get 15

Vendors publish deprecation notices one model at a time, so the pattern only shows up once the dates are collected in one place. We keep that collection as a public feed, and it now covers 31 models from five makers — OpenAI, Anthropic, Google DeepMind, Mistral AI and Amazon — each with both a launch date and a shutdown date announced by its own maker. Sorted by the year they launched, the median time from release to retirement is 43.3 months for the 2023 cohort (three models), 18.4 months for 2024 (nine models) and 14.6 months for 2025 (eighteen models). The single 2026 model with a published end date, Gemini 3.1 Flash-Lite, gets exactly 12 months. Put differently: a model launched in 2025 is scheduled to be switched off roughly three times sooner than one launched in 2023. The extremes are wide. The longest-lived are OpenAI's originals — GPT-3.5 Turbo (43.8 months) and GPT-4 (43.3) — both of which carried the company through the period when there was nothing newer to move users to. The shortest is Mistral's Devstral 2, given 7.7 months from release to shutdown. Sora 2 is the shortest-lived flagship: 11.8 months between its launch on 30 September 2025 and the API shutdown set for 24 September 2026. By maker, the averages are 22.3 months for OpenAI (13 models), 17.2 for Amazon (1), 16.1 for Anthropic (8), 15.2 for Google DeepMind (5) and 14.2 for Mistral AI (4). OpenAI's lead is entirely an artefact of age: strip out its three 2023 models and the rest of its line-up sits in the same 15-to-20-month band as everyone else. One caveat matters more than the numbers, and no one else is going to state it for us: this sample only contains models whose retirement has already been announced. Models still running with no end date in sight are, by construction, absent — so the true average life of an AI model is longer than 18.3 months, and the recent cohorts are the most affected by that bias. What the figures do establish is narrower and still uncomfortable for anyone building on top of these APIs: when a maker does set an end date, the window it grants has been getting shorter every year.