Which Gemini model answers Google AI Mode? It's now a menu — 3.7 Flash joins Auto and Pro for paid subscribers

What happened

Two announcements, one day apart. On 2026-08-13 Google published its Gemini 3.7 Flash launch post, pitching the model as "our most intelligent workhorse model yet for coding and agents," with introductory API pricing of $0.75 per million input tokens and $3.75 per million output through the end of 2026. That post says nothing about Search — we read it in full. The Search rollout came separately from the Search team: VP of Product Robby Stein announced on X that Gemini 3.7 Flash is rolling out in AI Mode for Google AI Pro and Ultra subscribers, globally in English, selectable from the model menu behind the "+" icon, where it sits under "Gemini 3 models" next to "Auto" and "Pro" (Search Engine Land and Search Engine Journal both covered the announcement by 2026-08-14; Search Engine Land expects a wider rollout to free users "in the coming weeks or months," which Google has not confirmed).

AI Mode's answer now depends on who is asking

This is the third generator change under Google's AI search surfaces in four months — Search Engine Journal notes Gemini 3.5 Flash became AI Mode's global default in May, and on July 21 Google said plainly that Gemini 3.5 Flash-Lite was "also rolling out in Google Search". What is new this time is not the swap but the menu: the generator behind an AI Mode answer now varies by subscription tier and by a user choice invisible to everyone else. A free user gets the default; a paying subscriber may be asking the newest model Google ships. And paid-first rollouts put the newest generator in front of exactly the users most likely to be doing serious vendor research. Meanwhile, the two things measurement tools can easily sample — the free tier and the Gemini API — are now each one generator among several, and Search Console reports AI Mode queries only in aggregate, with no per-model split.

Why this matters if you care about AI visibility

• A visibility baseline without a generator label is now ambiguous by construction. When we covered the July Flash-Lite rollout we made the one-generator version of this point: a measurement taken before a model swap belongs to a different generator than the one answering today. The menu adds a second axis — measurements taken on the same day can come from different generators depending on tier and selection. Date-stamp your measurements, record which surface and model produced them, and re-measure after rollouts like this one.
• The paid tier is a measurement blind spot. Rank-tracker-style AEO tools overwhelmingly sample free surfaces or APIs. What a Google AI Pro subscriber sees after picking 3.7 Flash from the "+" menu is visible only to subscribers who picked it — no public dashboard reports it, and Search Console's AI Mode data carries no model dimension.
• Model swaps shuffle who gets named — that is the one effect of this you can check on your own business. A generator change alters how retrieved sources get used and which names make the answer, even when the underlying index is unchanged. The practical move is unchanged from July: check what AI engines say when asked about your business the way a customer would ask, note the date, and treat any pre-rollout reading as belonging to the old generator.