Gemini 3.5 & Omni: What Google I/O Actually Showed

There's a particular genre of tech demo that Google has perfected over the years: warm, slightly caffeinated, staged in a way that feels spontaneous but definitely isn't. The IO 2026 session on Gemini 3.5 and Omni fits the mold—and that's not entirely a criticism. Sometimes a format becomes a format because it works.

What developer advocates Dave Elliott, Lavi Nigam, and Katie Nguyen (credited as Katie Winn in the session) showed off at IO was a two-headed announcement: a new text-and-code model in Gemini 3.5 Flash, and a generative video model called Gemini Omni that can take basically any input modality and turn it into video. The demos were live—or live-adjacent—and they mostly held together. The ideas underneath them deserve more scrutiny than a 25-minute fireside chat allows.

The Cost Argument, Examined

The central pitch for Gemini 3.5 Flash is a value proposition: near state-of-the-art results at a meaningfully lower price point. Elliott was careful with language here—"near state-of-the-art results," not performance—and Nigam backed it up with a reference to Sundar Pichai's keynote math: at scale, the efficiency gains translate to billions of dollars in savings for large enterprises.

The honest asterisk: Flash's per-token pricing has actually gone up from previous Flash versions. The argument is that you're getting so much more intelligence per token that the effective cost-per-useful-output is lower. That's a reasonable frame, but it's also one that requires enterprises to do their own math rather than trust Google's math. Any vendor can tell you their product is efficient at scale. The proof is in your specific workload, your token volumes, your tolerance for "near" state-of-the-art when your use case actually requires state-of-the-art.

The benchmarks they walked through—SWE-Bench for coding, MCP Atlas for tool-calling, terminal-level batch task performance—show genuine improvements over Gemini 3.1 Pro and Gemini 3.0 Flash. Nigam was upfront that the Pro model for this generation hasn't been announced yet, which makes the Flash-to-Pro-predecessor comparison feel slightly convenient. Still, a Flash model posting SWE-Bench results that compete with prior-gen Pro is not nothing.

What "Agentic" Actually Means Here

The word "agentic" gets thrown around so freely in AI marketing that it's started to lose resolution. In this context, it refers to a model's ability to call external tools reliably, chain those calls across multi-step tasks, and maintain coherent context through the whole sequence—the kind of thing where previous models would hallucinate a tool call or lose track of earlier state.

The improvements Nigam highlighted—MCP Atlas scores jumping from 78 to 83.6, tool-calling accuracy up to 56%—matter specifically because agents fail in compound ways. A 5% improvement in single-step accuracy becomes a much larger reliability gap when you're chaining ten steps. That's the math that makes benchmark improvements in this category more meaningful than they look on a slide.

The Gemini Enterprise Agent Platform context here is worth holding onto: Google isn't just selling a model, they're selling a lifecycle. The "chocolate bar diagram" Elliott referenced—build, scale, govern, optimize—is the framing they're using to position ADK (Agent Development Kit) and Antigravity CLI as the on-ramp, with agent registry, runtime, security, and evaluation as the things that keep enterprise deployments from going sideways six months in.

Nigam made an observation that felt like the most honest thing said in the session: "You can build agents, you can play with them, but unless they're evaluated and the organization sort of agrees and aligns with that, there's no point of doing that." He also flagged that evaluations have a shelf life—"six months from now, the evaluation you did on day one may not be relevant because you have a new vendor that used a different language." That's a real operational problem, and it's refreshing to hear it named out loud at a product launch instead of buried in the documentation.

Omni: The More Interesting Announcement

If Gemini 3.5 Flash is an evolutionary model story, Omni is the one that opens more questions.

The concept: a video generation model that accepts any input modality—images, text, existing video, audio, storyboards—and synthesizes them into coherent video output up to 10 seconds long. It's currently available on consumer surfaces (Flow, the Gemini app, YouTube Shorts) with API access coming to developers and Agent Platform.

Nguyen's demos were genuinely illustrative. She showed a generated video of a suitcase she called the "Omni case"—prompted to float and spin, which defies physics, but the model correctly applied realistic physics to the opening mechanism and rendered custom text/logo elements with high fidelity to the prompt. More interesting: running the same imperfect, grammatically informal prompt repeatedly produced consistent output. The Omni case logo rendered the same way each time. That kind of prompt-to-output stability is a real differentiator for production use cases; if a marketing team is generating product videos at scale, consistency isn't optional.

The storyboard-as-input capability is the one I find myself turning over. Nguyen showed a Nano Banana-generated storyboard—complete with text annotations—fed as a reference image into Omni, which then used it to plan and generate video with scene structure intact. The model treated the storyboard's written directions as intent, not just visual reference. That's a meaningful compression of the production pipeline for anyone who's sat through the gap between concept and execution in video production.

Elliott's photo booth anecdote gets at the underlying capability clearly: he walked into an IO photo booth wearing a shirt with a logo, asked to be shown presenting at IO, and the output put him on stage presenting a session on ADK—correctly inferring that ADK was contextually appropriate for the shirt's branding without ADK being spelled out anywhere. "That world understanding now comes to video," he said. The phrase is a bit of a hand-wave, but the demo is concrete.

The Questions This Raises

None of this is criticism for its own sake, but: Google is not a neutral party in AI infrastructure. They sell the models, they sell the platform those models run on, they sell the cloud compute those agents deploy to. The "chocolate bar diagram" of build-scale-govern-optimize is also, from another angle, a map of increasing lock-in. Every layer you adopt—ADK for scaffolding, Antigravity CLI for development, Agent Runtime for deployment, Agent Registry for management, Agent Evaluation for governance—is another layer that makes migrating to a different stack more expensive.

That's not inherently bad, and it's not unique to Google. AWS, Azure, and every major cloud player has been drawing the same map for years. But it's worth naming when a session opens with "how easy it is for you to get started with that journey." Easy to start doesn't always mean easy to leave.

The consistency and context-retention capabilities in Omni also open the deepest box in generative media: likeness. Nguyen showed her own face in a generated beach video, consistent frame to frame, her dog's features stable throughout. "Your face doesn't change," Elliott noted approvingly. That's a genuinely useful creative capability and also, plainly, the same capability that makes non-consensual deepfakes tractable to produce. Google's guardrails weren't discussed in this session; they presumably exist, but a launch demo is not where you stress-test them.

And then there's the evaluation problem Nigam raised, which doesn't fully get resolved by having an evaluation tool. Automated evaluations can tell you whether an agent performs consistently against a defined rubric; they can't tell you whether the rubric is right, or whether the right rubric today is still the right rubric after an org restructure. "The organization sort of agrees and aligns with that" is doing a lot of work in that sentence, and human alignment on AI agent behavior in enterprise contexts is genuinely hard in ways that no CLI makes easy.

What's visible at IO 2026 is a Google that has moved past the "we have models too" phase and is selling a complete production story—from idea to deployment to governance to ongoing evaluation. Whether that story holds in practice, at the scale and specificity of real enterprise workloads, is a question no demo can answer.

---

Marcus Chen-Ramirez is a senior technology correspondent at Buzzrag. He covered software infrastructure before he covered the people building it.