New AI model releases, May 25 2026: the dated event was a shutdown, not a launch

What the back half of May 2026 actually looked like

People type a precise date into a search box because they expect a precise launch behind it. For May 25, 2026 there isn't one. The frontier labs were between events. The only thing in the model world stamped with that exact date is a retirement. Here is the dated log for the window, with each entry traceable to a public source.

Sources: the GA of Gemini 3.5 Flash on May 19 is on Google's announcement and was covered at I/O 2026. The stable gemini-3.1-flash-lite release (May 7) and the gemini-3.1-flash-lite-preview shutdown (deprecated May 11, withdrawn May 25) are listed on Google's deprecations page.

Gemini 3.5 Flash was the real launch in the window

If you came here for a model to try, the answer is the one that shipped six days earlier. Gemini 3.5 Flash went GA on May 19, 2026 at Google I/O. Google put it at $1.50 input and $9 output per million tokens with a 1M-token context window, and pitched it as roughly 4x faster than comparable frontier models on agentic and coding work. It is the model worth wiring up; the preview that died on the 25th is the one worth removing from your config.

Gemini 3.5 Pro is not part of this window. At I/O, Sundar Pichai told the room to give them until the following month, which puts it in June 2026 with no firm date. So the late-May story is one GA (Flash), one stable graduation (Flash-Lite), and one shutdown (the Flash-Lite preview). That is the whole list.

The part nobody writes down: a retirement breaks more than a launch

A launch is opt-in. You read about Gemini 3.5 Flash, decide whether to adopt it, and nothing changes if you ignore it. A deprecation is the opposite. On May 25, every request to gemini-3.1-flash-lite-preview stopped resolving whether you were paying attention or not. If a script hardcoded that string, it started erroring. If a desktop app shipped a fixed list of selectable models, it kept offering a dead option until someone pushed an update and you reinstalled.

That is the gap in every "what released this month" piece. They optimize for the exciting half of the lifecycle and ignore the half that actually touches your code. The question that matters on a date like May 25 is not "what can I try" but "did anything I depend on just go away," and the answer this time was yes.

How fazm treats a model list as live data, not a hardcoded menu

fazm is a native macOS app that wraps Claude Code and Codex through the Agent Client Protocol. The relevant design choice for a deprecation date is this: the model picker is never a list baked into the app. It is whatever the running agent reports. A single function in the bridge, emitModelsIfChanged, takes the SDK's reported availableModels, normalizes them, and forwards the set to the UI only when it differs from the last one emitted.

// acp-bridge/src/index.ts
function emitModelsIfChanged(availableModels) {
  const transformed = availableModels
    .filter(m => m.modelId !== "default")
    .map(normalizeModelId)   // strip [1m] variants, bump stale ids
    .filter(dedupeByModelId);
  const json = JSON.stringify(transformed);
  if (json === lastEmittedModelsJson) return;  // no-op unless the set changed
  lastEmittedModelsJson = json;
  send({ type: "models_available", models: transformed });
}

The consequence is symmetric. When a lab adds a model, it appears in the picker without an app update. When a lab withdraws one, it falls out of the reported set and disappears from the picker. There is no version of fazm that pins gemini-3.1-flash-lite-preview in a place that can outlive the endpoint. A retirement date is, for the UI, just the moment one row stops being sent.

Switching to whatever did survive, or to a model that just launched, is a per-chat action rather than a reinstall. fazm calls session/set_model on the active session (see acp-bridge/src/codex-query.ts), so you pick the new option from the menu and keep working in the same window with full history intact. If the model you want lives behind your own gateway, fazm also accepts a custom Anthropic-compatible endpoint, so a freshly released model exposed through a proxy is reachable the day it appears, not the day an app update ships.

Questions people actually ask about May 25, 2026