How I AI: My Live Test of Google I/O’s New AI Tools—From Gemini 3.5 Flash to Omni Video

Google announced a flood of AI products and updates at I/O 2026. I opened the tools on launch day and tested what actually worked across Antigravity, AI Studio, Gemini image and video generation, Flow, Pomelli, and Stitch.

Instead of treating the keynote demos as proof, I checked what was available in my own accounts, how fast the workflows felt, and whether the generated outputs were usable without cleanup.

The results were uneven in a very real launch-day way. Antigravity successfully turned an internal blog tool into an authenticated API. Omni generated a surprisingly good animated short from a child’s drawing. Meanwhile, the promised Workspace integration never appeared in my account, the avatar flow failed after scanning my face, and Gemini’s portrait editor produced someone who looked nothing like me.

Developer tools that shipped with real utility

In this episode of How I AI, the developer story centered on Gemini 3.5 Flash and Google’s broader push into agent workflows. A lot of the concepts looked familiar if you already use Codex or Claude Code: projects, scheduled tasks, subagents, CLI workflows, and long-running goals. The question was whether Google’s speed claims and multimodal strengths changed the day-to-day experience enough to matter.

Turning a blog generator into an agent-ready API

Google described Gemini 3.5 Flash as a fast, highly capable coding model for agentic workflows. Rather than rely on benchmark charts, I tested it directly inside Antigravity 2.0 using a real internal task from ChatPRD.

The starting point was ChatPRD’s existing admin blog generator, which already converted podcast videos into draft blog content through a web interface.

The goal was simple and practical: replace the manual UI with an authenticated API endpoint an AI agent could call programmatically.

The workflow itself was intentionally lightweight so the model had to infer structure from the repository and existing implementation patterns:

1. 1. Open the ChatPRD repository inside Antigravity 2.0 and scope the project workspace.
2. 2. Give Gemini 3.5 Flash a short natural-language implementation request instead of a detailed spec.

We have a blog generator UI at, admin tools. I want to turn this into an API that an AI agent can use instead of a web UI for our team. Please build.

1. 3. Approve repository access, allow the agent to inspect files, and let it generate the API implementation and supporting documentation.

The result was functional. Antigravity generated an API-key authenticated endpoint that triggered the existing blog-generation workflow and supported passing assets like featured-image URLs. It also generated documentation explaining how the endpoint worked. I did not personally verify Google’s broader claims that the model is dramatically faster than competing coding models, but the task completed cleanly and without much intervention.

One feature I want to test more is the new /grill-me slash command. The idea is that the agent aggressively clarifies vague requirements before building, similar to the clarification flows in Claude Code but with a more forceful product framing. I had not used it enough yet to know whether it meaningfully improves outcomes or is mostly branding.

Testing the promised Workspace connectors in AI Studio

Google AI Studio also announced built-in access to Workspace data sources including Gmail, Sheets, Drive, and Calendar. The pitch is obvious: instead of wiring up MCP connectors or custom integrations, you could build lightweight internal tools directly against your Google account.

I tried a very normal personal productivity use case: generating a month view of family weekend events and sports schedules from my Google Calendar.

1. 1. Open AI Studio using the Google account tied to the calendar data.
2. 2. Prompt AI Studio to create a planning app around upcoming family sports events stored in Google Calendar.

Make me an app to manage the next month of weekend events with my kids, in particular, our sporting events. They are all on my personal calendar in Google.

I could not access the Workspace integration at all. I checked settings, connectors, and available tools, but the feature simply was not exposed in that account despite being announced publicly. That does not mean the integration was universally broken, but it did reinforce a recurring theme from launch day: availability mattered more than the keynote narrative.

Creative tools that felt promising and unfinished at the same time

The second half of this episode of How I AI moved into Google’s creative stack: Gemini image generation, Omni video, Flow avatars, Pomelli branding tools, and Stitch for interface design. The products overlap heavily, the naming is confusing, and several workflows still felt experimental, but a few stood out.

Trying to generate a polished podcast portrait

I tested Gemini’s updated image workflow by uploading a screenshot from the podcast and asking the model to upscale the image, preserve my face, and replace the background with a professional podcast studio.

The requirement was not complicated. I wanted a cleaner promotional portrait that still looked recognizably like me.

1. 1. Upload a podcast screenshot into Gemini image generation.
2. 2. Prompt the model to upscale the image, beautify the portrait, preserve identity, and generate a studio-style background.

upscale and beautify because I wanna be pretty. The image of this podcast host. Change the background to a professional podcasting studio.

The generated image was polished in a generic AI-photo sense, but the identity preservation failed badly. The output looked like a different person entirely. The rendering quality itself was reasonably photorealistic, and the model generated the image quickly, but for creator workflows where likeness matters, this specific test was unusable.

Animating a child’s drawing into a short film

The strongest demo of the episode came from Omni video generation inside Gemini. Google has been leaning heavily into multimodal workflows, and this was the first place where the combination of image grounding and video generation genuinely felt compelling.

I used a hand-drawn superhero sketch from my child and asked Omni to turn it into a short classroom escape story.

1. 1. Take a photo of the drawing and upload it into Gemini’s video workflow.
2. 2. Use the uploaded image as the visual reference for the generated scene.
3. 3. Prompt the model with a simple one-sentence action sequence describing the superhero breaking a kid out of class to go have fun.

animate this superhero, breaking a kid out of class to go. Have fun

The resulting ten-second clip kept the original character recognizable while animating the classroom sequence in a coherent way. That consistency matters. A lot of image-to-video systems drift away from the source material almost immediately, especially with rough hand-drawn references. In this case, the grounding held up well enough that my kid immediately recognized the character.

The avatar workflow that failed after the face scan

Flow is Google’s more production-oriented video environment. Beyond scene editing and cinematic controls, it includes a guided avatar system designed to create reusable digital characters from a phone-based face scan.

I tried to create a reusable avatar version of myself for future video scenes.

1. 1. Open Flow and choose the Create an avatar workflow.
2. 2. Scan the QR code, continue the capture flow on a phone, and grant camera access.
3. 3. Follow the guided prompts by reading numbers aloud and rotating your head for the facial capture process.

The upload completed, but the avatar never appeared. The workflow captured my face data, processed the upload, and then effectively stalled. This was one of several moments where the ambition of the launch exceeded the reliability of the product experience. A later episode eventually completed the same flow successfully, but on launch day it failed end to end in my account.

Using Pomelli for branding and Stitch for UI generation

Pomelli and Stitch are aimed at adjacent creative workflows. Pomelli focuses on brand extraction and marketing assets. Stitch focuses on interface generation and editing.

I tested both tools using ChatPRD as the source material. The goal was to extract the existing brand identity into generated assets and then create a Google I/O-inspired application screen.

• Pomelli analyzed the ChatPRD site, extracted the colors, copy, and general brand direction, and generated a working landing page. The result was coherent enough to demonstrate the workflow, but it still had the polished-generic feel common to many AI-generated marketing pages. I would use it as a rough first draft, not as production-ready design.
• Stitch was more impressive. I referenced the Google I/O site and asked Stitch to generate a related app interface. The system streamed the design directly into the canvas while building the layout and visual system in real time. Even though I accidentally selected mobile instead of web, the interaction model itself felt fast and tangible in a way that matched Google’s claims about Flash-powered generation.

What I would actually copy from these workflows

Three workflows stood above the rest in practical value. Antigravity handled a constrained coding task with minimal prompting and produced inspectable output. Omni’s reference-based video generation showed real promise for turning sketches, concepts, and source imagery into coherent clips. Stitch made rapid UI exploration feel much faster than traditional blank-canvas design work.

The weaker results mostly came down to reliability and availability. AI Studio’s Workspace integrations were inaccessible in my account. Flow’s avatar workflow failed after the upload step. Gemini’s portrait editing could not preserve identity well enough for creator use. Those are exactly the kinds of details that determine whether a tool becomes part of a real workflow or stays a conference demo.

If I were adopting these tools today, I would start with tightly scoped coding tasks, grounded image-to-video generation, and fast interface prototyping. Those workflows already produce artifacts you can inspect, iterate on, and decide whether to ship. The more ambitious agent and avatar workflows still require patience, fallback plans, and human judgment when the automation breaks.

The real lesson from launch day

• Availability and execution quality mattered more than the announcement itself. A smaller feature that reliably generated a usable artifact was more valuable than a headline capability that was unavailable, hidden behind rollout gates, or unstable in practice.
• The next meaningful test is not whether these products can demo well. It is whether the missing integrations, unreliable flows, and inconsistent outputs become dependable enough to trust in repeated real-world use without lowering the acceptance bar.