Google published its annual AI recap on December 22, listing 60 announcements the company made throughout 2025. The post, organized month-by-month, covers updates to Gemini, NotebookLM, Pixel hardware, and the expansion of AI into Google Search. CEO Sundar Pichai's name appears early, attached to his February claim that AI represents "the most profound shift in our lifetimes."
The Gemini 3 bet
November's Gemini 3 announcement was the year's marquee moment. Google unveiled the model on November 18 and, in a departure from its usual staggered rollouts, pushed it across Search, the Gemini app, AI Studio, and Vertex AI on day one. The company claimed 2 billion monthly users gained instant access through AI Mode and AI Overviews.
The numbers behind Gemini 3 are genuinely impressive if you take them at face value. Google says Gemini 3 Pro hit a 1501 Elo score on the LMArena leaderboard, becoming the first model to cross 1500. Demis Hassabis, who runs DeepMind, said previous models would "lose their train of thought" around step 5 or 6 of complex reasoning chains, while Gemini 3 reliably handles 10 to 15 steps.
Whether users actually notice the difference in daily use is another question. The Hard Fork podcast's Kevin Roose captured the skepticism nicely: Google "struggled in AI for a couple of years there" with Bard and early Gemini versions, so "the question is: is this them taking their crown back?" The recap doesn't engage with that uncertainty at all.
By mid-December, Google had already shipped Gemini 3 Flash, the speed-optimized variant. The company says it's processing over 1 trillion tokens per day through the API. Gemini 3 Flash now powers AI Mode globally, which raises an obvious follow-up: how much of that usage is people actually finding it useful versus Google simply making it the default?
Search gets rebuilt around AI
The transformation of Google Search was arguably the bigger story. AI Mode launched in March as an experimental feature for power users who wanted AI responses for more of their queries. By December, Nick Fox confirmed it had reached 75 million daily active users worldwide.
The feature uses what Google calls a "query fan-out technique," breaking your question into subtopics and running multiple searches simultaneously. For something like comparing sleep trackers across different form factors, it can synthesize results in ways a traditional search couldn't. That's the pitch, anyway.
What's missing from the recap is any acknowledgment of the delays. At I/O in May, Google showed personal context features that would connect AI Mode to Gmail and other apps, letting it tailor results based on your email confirmations and calendar entries. None of that has shipped publicly. Fox told the AI Inside podcast these features are "still to come" with no timeline.
The recap mentions AI Mode queries run two to three times longer than traditional searches. Google frames this as users asking more complex questions. Critics might note it also means more time on Google's properties and more opportunity to serve ads.
Flow and the filmmaking pitch
Google I/O in May brought Flow, an AI filmmaking tool built on Veo 3. The company partnered with director Darren Aronofsky's studio Primordial Soup for the launch, lending some Hollywood credibility to a product that otherwise might have looked like another AI toy.
Veo 3 adds native audio generation, meaning it can create sound effects, ambient noise, and even dialogue that syncs with lip movements. Combined with Imagen 4 for still images and Gemini for natural language prompting, Flow represents Google's attempt at a complete creative production pipeline.
The pricing tells you who Google thinks will actually use it. Flow requires a Google AI Pro subscription for 100 generations per month, or Google AI Ultra at $249.99 monthly for native audio and higher limits. That's not hobbyist pricing.
By July, Media Matters reported racist and antisemitic videos generated with Veo 3 were appearing on TikTok. Ars Technica noted the model can't reliably detect when users are prompting for content that exploits racist tropes through indirection. The recap, predictably, doesn't mention this.
NotebookLM keeps shipping
NotebookLM had a quieter but more consistent year. The research tool started 2025 with basic audio overview features and ended it running on Gemini 3 with capabilities that would have seemed excessive 12 months ago.
The November updates were particularly dense: Slide Decks and Infographics powered by Nano Banana Pro, a "Thinking UX" that shows the model's reasoning process, Deep Research for autonomous web sourcing, Microsoft Word document support, and the ability to autogenerate Audio Overviews on source upload.
Data Tables arrived in December, letting users ask NotebookLM to scan documents and synthesize specific variables into structured tables exportable to Google Sheets. For researchers dealing with qualitative data across multiple sources, that's genuinely useful rather than just impressive.
The tool now supports 35+ languages across 180+ regions. Google Workspace customers can access NotebookLM Plus through various education and enterprise add-ons, though the pricing structure has gotten complicated enough that Google published separate documentation just to explain who qualifies for what.
Hardware as AI showcase
The Pixel 10 series launched in August with the Tensor G5 chip, which Google designed with DeepMind specifically for on-device AI. The phones ship with Gemini Nano, running complex generative AI tasks locally rather than in the cloud.
Magic Cue connects data across apps to surface relevant information proactively. Text a friend asking about your flight arrival and it pulls details from your itinerary without you opening a separate app. That's the demo, at least.
Camera Coach uses Gemini to offer real-time photography suggestions. Voice Translate handles call translation. Pixel Journal competes directly with Apple's journaling app.
The Pixel 10 Pro Fold added IP68 water and dust resistance, a first for Google's foldables. At $1,799, it's priced competitively with Samsung's folding phones while offering Google's full AI stack.
What the hardware announcements really demonstrate is Google's distribution advantage. The company doesn't need Pixel to dominate smartphone sales. With Android on over 3 billion devices, the goal is showcasing capabilities that Samsung, Xiaomi, and others will adopt, spreading Gemini to users who never buy a Google phone.
