
Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Google’s real flex was Gemini 3.5 Flash, not just the flashy demos — Alex says the whole keynote rested on a new Gemini 3.5 Flash model that’s “four times as fast” as other frontier models, beats Google’s previous Pro model on agentic benchmarks, and even showed up in pricing talk around $9 per million tokens with a 1M-token context window.
The biggest crowd curiosity was Omni: Google’s new video-native, chat-driven editing model — Multiple creators compared Omni to “nano banana for video,” with early testers saying it can edit clips conversationally, pull in world knowledge, preserve characters surprisingly well, and is already available to some Gemini users through the chat interface.
Google is pushing AI from cloud to device with Gemma, and it’s moving fast — Gemma lead Olivia says Google’s open model is built for phones, laptops, multilingual use, reasoning, function calling, and low-spec environments, while Alex notes Gemma crossed 100 million downloads in roughly six weeks.
Google Search quietly got one of the weirdest demos of the day: generated UI in the search box — Jane Manchun Wong called out the new ability to generate custom interfaces directly inside Search, not just text or code, and Alex immediately tried a leaked URL from Josh Woodward’s onstage demo to see it in the wild.
The glasses got buzz, but the room still wanted proof they’re real products, not another eternal demo — Greg Eisenberg said Google’s new audio-first XR glasses looked more elegant than Meta Ray-Bans, while longtime Google Glass diehard Alan Fennberg said the magic is there but third-party integrations are still the big unanswered question.
The stream itself captured the real I/O vibe: creators, PMs, and analysts all agreed this felt more iterative than shocking — Matt called the keynote “marginal upgrades,” but still singled out Omni and Flow as standout launches, while others framed I/O 2026 less as one giant reveal and more as Google stitching together releases it’s been shipping all year.
Ray opens from Hawaii while Alex roams Google HQ in Mountain View, very much doing the “we’ll do it live” version of I/O coverage. Alex is sweating through the creator-builder tent, sipping personalized coffee to cool his phone, and setting the scene: about 5,000 people, Logan Kilpatrick hosting creators from YouTube, TikTok, Instagram, X, and Substack, and a keynote that just unloaded “an incredible amount of stuff.”
Alex’s first big takeaway is that the real story is Gemini 3.5 Flash: faster, cheaper, and agentically stronger than Google’s prior Pro model. Matt from the creator crowd gives the first reality check of the day — he liked Omni and Flow, but says I/O felt lower-energy than past years, more “marginal upgrades” than earth-shaking reveals, partly because Google now ships product updates all year instead of saving them for one stage moment.
Alex grabs Jane Manchun Wong, who says one of the most exciting moments wasn’t a product feature at all, but Google talking about a “clinical stage” cancer effort. She also highlights the generative UI demo in Search — Google can now generate custom interfaces, not just text answers, directly inside Search — and Alex admits he instantly did a very Jane thing: screenshotting Josh Woodward’s onstage URL and trying it himself.
Olivia, who leads Gemma, frames it as Google’s open model for “local AI,” built to run on constrained environments like phones and laptops while still handling reasoning, function calling, multimodal work, and multilingual use. She’s careful not to oversell on-device agents just yet, saying agent harnesses are still mostly server-side today, but Alex keeps pushing the obvious angle: if people are nervous about sending private data to the cloud, Gemma starts looking like Google’s answer.
Amar Reshi from AI Studio says the builder side of the launch is just as important: prompt-to-Android-app generation inside AI Studio, with Play Store distribution eventually coming straight from the product, plus a mobile AI Studio app on the way. He’s personally most hyped about Omni, especially for editing, because video edits require absurd precision and the promise here is “nano banana style” conversational control over video instead of the usual pain.
Analyst Max Weinbach says he was already playing with “anti-gravity” during the keynote and saw it evolving from an IDE into more of an agent manager. He zooms in on the infra story — TPU v8e, inference economics, and Google’s ability to serve a giant model fast and cheap — while also dropping the line Alex clearly loves: Google basically saying “your margin is my opportunity.” Others, like Greg Eisenberg, are impressed by Google’s safer consumer-agent framing with Spark, but still notice what’s missing, like MCP integrations on day one.
The line for Google’s new glasses is long and sweaty, and Alex eventually bails on waiting, but not before capturing the funniest possible detail: a longtime attendee still wearing 13-year-old Google Glass. Alan Fennberg says the new experience finally feels like the natural-language, always-available assistant people imagined back then, but he’s still waiting to see how developers plug in their own tools and workflows — classic I/O optimism with one eyebrow raised.
As Alex keeps roaming, the human side of I/O takes over: creators from Spain and Japan, niche AI podcasters, ad-makers, and Two Minute Papers’ Károly Zsolnai-Fehér all drift into frame. The best bit is how unfiltered it all feels — people are sunburned, half in line, half between demos, casually name-dropping Jeff Dean and Demis Hassabis, and still trying to answer the same basic question: which of these launches is actually ready, and which is just a really good trailer?
Share
Keep Reading
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.

Playbook
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.

Playbook
Learn how tasteful prompting helps you move beyond generic AI output by shaping context, style, and judgment from the start.

Playbook
OpenAI shipped /goal for the Codex CLI. It turns a prompt into a persisted, self-continuing contract.