Gemini Pre-I/O Rumor Roundup: Spark, Omni, Cappuccino Two Days Out

Two days out from the Google I/O 2026 keynote. May 19, 10:00 AM PT, Sundar Pichai opens at Mountain View's Shoreline Amphitheatre, and what exactly drops there has been a moving target for weeks. The candidate set narrowed fast over the past week. 9to5Google, Gadgets360, and Knightli each peeled off a separate signal, and on X a single user's 'Spark / Omni / Veo / 3.5' roundup pulled 7,200 impressions on its own — effectively the consensus D-2 scenario now. The surviving leaks are three: a 24/7 background agent codenamed Spark, a video model called Omni, and a next-gen LLM checkpoint dubbed Cappuccino.

Spark — A 24/7 Agent That Reads Your Gmail

9to5Google's May 14 report describes Spark as something other than a chatbot. It is a separate agent hidden inside the Google app v17.23 binary, engineered to run in the background without a prompt, stitching context across Gmail, Calendar, open tabs, and apps.

A May 16 X thread put it bluntly — 'reads your Gmail, watches your tasks, checks your calendar, can take actions in the background' — and pulled 1,682 impressions and 13 likes. Another user in the same wave dropped a second handle: 'Gemini Spark = Remy,' an internal codename. Code leaks point to a Skills system and a task scheduler as distinct components of the same agent.

Omni — 790K Views in Half a Day, the Video 'Nano Banana Moment'

Thread two is the video model, Omni. Gadgets360 reports it as a unified Gemini-based video generation, editing, and remix model under internal testing ahead of I/O 2026. Separate line from Veo, but sharing the video stack.

The scale of the signal showed in the first demo clip that leaked on X early May 17. A Dutch-language post wrote: 'The first demo of Omni leaked today. 792k views on X in a few hours. The text inside the video is just correct — no garbled letters. For images Nano Banana did this. For video, this is it.' Another comment pushed back gently — 'Not Veo 4. The 9-second length alone tells you it's a different line.' Runway, whose $5.3 billion valuation rides on owning AI video, had already given a pre-emptive interview to TechCrunch — the timing reads less like a coincidence and more like a defense against I/O Monday.

Cappuccino — Rename Rumor and the Game-Build Claim

Thread three is the loudest. A May 15 X post claimed: 'Google renamed Gemini 3.2 to 3.5. Leaked checkpoint Cappuccino is already building full playable games from scratch.' Same day, Knightli.com followed up framing Cappuccino as the internal codename for an unreleased Gemini 3.5 Pro.

3.2 versus 3.5 was already the central question of last week's roundup. What changed this week is a codename attaching itself to the higher-number side. The AI Studio A/B exposures, the brief 'Gemini 3.5 Pro' option in Antigravity settings, and now the single identifier Cappuccino — three traces in the same direction. The 'builds full playable games' line came as text description without an accompanying demo video, though, so until a live walkthrough lands on stage it is one notch softer than the other two threads.

Veo Upgrade, Googlebook, XR Glasses — Adjacent Signals

If those three are the main stage, three more sit in the wings. Veo is the established video axis — the I/O question is how much Omni absorbs and where Veo 4 picks up the rest. The 'Googlebook' AI laptop family is in the same orbit. A May 17 X post summarized: 'I/O 2026 targets actionable AI. Confirmed: Googlebook AI laptops and cross-app Gemini integration. Rumored: Gemini 3.5 Pro, Spark, Veo upgrade, Omni.'

XR glasses surface briefly. One multilingual news roundup grouped 'agentic Android AI, seven new voice models, Omni video generation, XR smart glasses' in a single line. The glasses signal is the thinnest of the bunch — no demo, no code leak. A brief cameo on stage would already count as success at this point.

What Takes the Stage, What Stays in the Rumor Pile

The D-Day checklist is short. First, does Spark ship as its own surface, or get folded into the Gemini app as a new mode? Second, does Omni get a live keynote demo, or does it land via a blog post without one? Third, does the next-gen text model arrive as Gemini 3.5 Pro on stage, or does it drop to 3.2 and turn the rename rumor into a dead lead?

On the adjacent side, the Veo upgrade is close to a lock; Googlebook's interesting question is form factor and pricing, not whether it gets announced. The XR glasses are the most likely to slip out of this keynote entirely and get pushed to I/O 2027. That the rumor pile has narrowed to three core threads two days out is itself the story: the audience the keynote can land on has gotten that much wider.