Google I/O is now eight days away — May 19-20, 2026[1]. Based on the leaks, the developer-relations pre-briefings, and the patterns Google has shown across Gemini launches, here is what to expect on May 19 — and which announcements actually matter.
The headline: Gemini 4
Gemini 4 is the most-discussed pre-announcement. Expectations are:
- Native image and video generation in a single model. Current Gemini generations split between Imagen for stills, Veo for video, and Gemini for text. Gemini 4 reportedly does all three in one call.
- Faster response times and improved reasoning. Specifically, a reduction in the median latency from query to first token, and improvements on long-context reasoning tasks where Gemini 2.5 already led the market.
- Deeper integration across Google services — Workspace, Search, Maps, Photos. The model has access by default; you do not need to grant per-service.
Whether Gemini 4 leapfrogs Claude Opus 4.7 and GPT-5.5 Instant on the standard benchmarks is the open question. Google's last several Gemini launches have been incremental on benchmarks but strong on integration. Expect more of the same — Gemini 4 will win on Workspace integration, native multimodal, and long-context, while Claude and GPT-5.5 stay ahead on hard reasoning.
Proactive Assistance — the genuinely new feature
The most interesting expected announcement is Proactive Assistance[2][3]. The pitch:
- Gemini watches your Gmail, Calendar, Drive, contacts, text messages, and what's on your screen.
- It surfaces suggestions at the right moment — "you have a meeting in 10 minutes, here is the brief", "this email needs a reply, here is a draft", "you forgot to ship this PR for the deadline tomorrow".
- The integration is system-wide on Android, not app-by-app.
If Google ships Proactive Assistance well, it is the closest mainstream consumer-AI product to a real "Jarvis". The privacy concerns are predictable — Google watching everything on your phone is a lot to opt into — but the productivity payoff for users who do opt in is substantial.
The competitive read: Apple Intelligence is supposed to ship a similar feature this autumn. Google getting there first puts them ahead on the platform-AI race.
'Remy' — the agentic Gemini
Less confirmed but heavily rumoured is Remy, a fully agentic version of Gemini[4]. Described as a 24/7 AI agent that runs digital errands, monitors routines, and uses connected apps.
This is Google's answer to Manus and OpenAI Operator. The question is whether Remy is genuinely competitive, or whether it lands a step behind the field.
Things that would make Remy compelling:
- Sub-agent parallelism (matching Manus's Wide Research)
- Cross-app workflows (sending emails, updating sheets, scheduling meetings as one task)
- Affordable pricing (bundled into Google AI Ultra, not a separate SKU)
Things that would make Remy a damp squib:
- Chrome-only (like Chrome Auto Browse, the Mariner successor)
- Single-agent architecture (already behind Manus)
- Premium pricing (over $50/month standalone)
I will reserve judgement until launch, but Google's track record on agentic products has been mixed.
Android 17 and Aluminium OS — two separate launches
Two distinct things commonly conflated in the pre-I/O reporting:
- Android 17 is the platform launch for phones — the usual API additions plus deeper Gemini integration via system services.
- Aluminium OS is a separate desktop operating system for PCs (HP, Lenovo, Acer, ASUS confirmed), built on top of Android 17 with a custom window manager, taskbar, and virtual desktops[3]. It is not a phone feature; it is Google's renewed push into Chromebook-replacement territory.
Expect both to be announced together at I/O. Don't confuse them. (I covered the Android 17 developer-side changes in a separate post.)
AI Ultra Lite — the new pricing tier
Confirmed earlier this month: Google is preparing an "AI Ultra Lite" tier[5]. Sits between $20 Pro and $250 Ultra. Reporting suggests the band is $50-$150/month; my guess is around $60-70 if Google wants it competitive with stacked ChatGPT Plus + Claude Pro.
For developers, the Ultra Lite tier likely includes Gemini 4 Pro API access at preferential rates — making it the natural choice for engineers who use Google's models heavily but cannot expense the full Ultra plan.
What I am actually watching for
Three things, in priority order:
- Gemini 4 long-context benchmarks. If Gemini 4 ships with 5M+ context at competitive accuracy, it is the model to use for document-heavy work.
- Proactive Assistance launch markets. If it launches US-only first, the EU AI Act compliance work is non-trivial. Watch for whether Google does a US-only soft launch or a global launch.
- Remy architecture details. If it is multi-agent, the autonomous-agent race tightens. If it is single-agent, Manus remains the lead platform for the rest of 2026.
The verdict
Google I/O 2026 is the most consequential Google AI event since Gemini's original launch. Expect a lot, calibrate expectations downward on agentic capabilities, and watch the long-context and Workspace-integration claims closely — those are where Google is most likely to deliver something genuinely differentiated.
Eight days. The whole industry will be watching.