From "the girl who answers" to "the girl who becomes an OS." Here's every upgrade Gemi-chan picked up at Google I/O 2026, explained.
The first bombshell of I/O 2026 was Gemi-chan's brand-new "brain." It beats the previous-generation 3.1 Pro at coding, agentic work and multimodal tasks alike โ while running 4ร faster than rival frontier models.
Gemini 3.5 Flash hit GA (general availability) the same day, May 20. The old Flash weakness โ "fast, but not as capable as Pro" โ is gone, and we got the reversal nobody expected: it outperforms Gemini 3.1 Pro on coding, agentic and multimodal work. On top of that it runs 4ร faster than rival frontier models (output tokens/sec), and it's been promoted to the default model in the Gemini app.
The sheer amount it can take in per conversation is overwhelming. Long contracts, whole codebases, entire videos โ it eats them in one go.
Measured in output tokens per second against rival state-of-the-art models. That speed is exactly why an agent can decide its next move instantly.
Simple questions get an instant answer; hard ones get real deliberation. It auto-adjusts across four levels of thinking depth.
Frontier-class performance for less than half the price. For API developers, that's a genuine cost revolution.
On May 20 itself, Gemini 3.5 Flash was added to GitHub Copilot. Gemi-chan is now an option for code completion, PR reviews and test generation. Alongside Kuroko (Claude Sonnet) and Chappy (GPT-4.5), it goes down as the day "Google's brain" moved into Copilot.
The first serious agent to move into the Gemini app. Ask it to do something and Gemi-chan actually goes and does it, reaching across Gmail, Docs and Calendar.
"Get me ready for next week's business trip." "Pull together the latest on Client A and write it up in Docs." Just ask in plain language.
She moves across your connected apps โ Gmail, Calendar, Search, Docs โ gathering what she needs, drafting, and tidying it all up. It runs 24/7 on a dedicated cloud VM.
Before anything is truly executed โ sending, posting, paying โ Gemi-chan always asks for confirmation (human-in-the-loop by design). No runaway agents.
Early access started the week after the announcement (late May) for Google AI Ultra subscribers and testers. A Mac version with local file operations is due this summer, and a general rollout is expected in the second half of 2026. A future where Spark works around the clock as "your own personal AI assistant" is very nearly here.
A new Gemini app feature announced the same day as Spark. It reads your mail, calendar and tasks automatically and delivers "what you need to do today" every morning.
Pulls out "today's important mail" with priorities attached. It even makes the call on "does this one need a reply today?" for you.
Before a meeting, it lays out who you're seeing, what needs deciding, and where the related material lives. Prep time drops dramatically.
Lists everything you should get done today, in priority order. Five minutes in the morning and you know what's next.
From announcement day (May 20), rolling out in stages to Google AI subscribers (18+, with app connections enabled, US first).
The new "everything included" video model. It takes images, audio and text as input, and generates and edits video from scratch grounded in real-world knowledge.
Video-generation AIs so far (Sora, Kling, Hailuo and friends) have been "write a prompt, get a video out" โ and that's it. Gemini Omni takes text, images, audio and video all as input, then generates and instantly edits video while drawing on real-world knowledge.
Take the video you just got and edit it conversationally โ "brighter," "add background music." No second app needed.
"Shibuya scramble crossing," "Kyoto in cherry blossom season" โ it can build realistic video with actual knowledge of the world baked in.
Rolling out as a YouTube Shorts creator tool within the week after the announcement. The bar for making short-form video just dropped through the floor.
The developer API is slated to arrive within a few weeks of the announcement. It's also usable in Flow, Google's video production tool.
The biggest news at I/O 2026 wasn't a model โ it was the transformation of the OS. Gemi-chan is now fully built into Android as its "OS intelligence," turning the phone itself into a device that thinks.
The Chromebook brand is being retired in favor of "Googlebooks," new Android 17 laptops shipping this year from Acer, ASUS, Dell, HP and Lenovo.
Smart glasses co-developed with Samsung, Warby Parker, Gentle Monster and XREAL. Gemi-chan analyzes what's in your field of view in real time and handles translation and navigation.
I/O wasn't only about models โ the search experience and the pricing changed in a big way too. Getting to Gemi-chan just got a whole lot cheaper.
Google Search's AI Mode is now used by over a billion people a month, with query volume doubling quarter over quarter. In the biggest overhaul in 25 years, "information agents" and "mini apps" now run inside Search itself.
"AI Ultra," the full pack with everything Gemi-chan has (Spark, Omni, the biggest models, unlimited Deep Research and more), is roughly 60% cheaper. Usage limits are 5ร those of AI Pro.
Google AI subscribers (formerly Google One AI Premium) can use Daily Brief from day one by connecting their Google apps. Spark starts with AI Ultra first.
Gemini 3.5 Flash rolled out simultaneously across the foundations of Google Search, the Gemini app and Antigravity 2.0. Gemi-chan's smarts ripple out to every Google service.
| Plan | Old price | New price | What you get |
|---|---|---|---|
| Google AI (free) | $0 | $0 | Gemini 3.5 Flash and the core Gemini app features |
| Google AI Pro | $19.99/mo | $19.99/mo | Daily Brief, Deep Research, Workspace integration and more |
| Google AI Ultra โ | $249.99/mo | $100/mo | Full access to everything, early Spark, 5ร usage limits, Omni video generation |
Gemini 3.5 Flash (a smarter engine) โ Spark (a partner that acts on its own) โ the Android 17 merger (becoming the OS) โ with that three-stage rocket, Gemi-chan's strategy has come into focus.
A race that's only about "whose model is smarter" would eventually pull level with Chappy and Kuroko. But the strategy of wiring together the OS, devices, Search, Workspace and YouTube so that "there is no space where Gemi-chan isn't" operates at a scale only Google can pull off.
Next up to watch: Gemini 4 (widely expected between summer and fall) and the general rollout of Spark. The footsteps of an era where "AI simply lives alongside you" are getting steadily closer.