Omni Flash Is Shipping — Not Just a Demo

Our previous coverage flagged Demis Hassabis invoking singularity language at Google I/O 2026. But while everyone argued about what that meant, Google quietly shipped something real.

Gemini Omni Flash is now live inside Google's AI video platform, Flow, according to blog.google. Not a waitlist. Not a preview. Available now.

Google has a long history of announcing AI capabilities that take months — sometimes a year — to actually reach users. Omni Flash skipped that lag.

What Omni Actually Is

According to blog.google, Gemini Omni is designed to take any input — photo, video, text — and produce any output. The long-term vision is a fully generalist media model. Right now, it starts with video.

Google says Omni combines Gemini's general knowledge with improved physics simulation — gravity, fluid dynamics, kinetic energy — to produce more realistic scenes. It also stamps every AI-generated video with SynthID, Google's imperceptible digital watermark, verifiable through the Gemini app, Chrome, and Search.

The watermarking detail is getting almost zero coverage. In an era of deepfakes and AI slop flooding the internet, it's a meaningful technical safeguard that deserves attention.

The Honest Review: Better, But Not Perfect

Allison Johnson at The Verge ran Omni Flash through its paces with a practical, repeatable test: recreating AI videos of a kid's stuffed deer — the same test she ran on the previous Veo model five months ago.

Her verdict: genuinely improved, still weird.

Omni handled character consistency better than Veo. Prompts translated more accurately to finished video. But the AI jump scares are still there — a stuffed deer inexplicably flipping orientation mid-skydive, the kind of uncanny glitch that reminds you this is still a probabilistic model guessing at physics, not actually understanding it.

The Verge's framing is fair: this is a real improvement, not a revolution. It tracks with what the benchmarks actually say.

Gemini 3.5 Flash: The Numbers That Got Ignored

Most mainstream coverage skipped this entirely.

According to blog.google, Gemini 3.5 Flash — also launched at I/O 2026 — outperforms Gemini 3.1 Pro on coding and agentic benchmarks. Specific numbers: Terminal-Bench 2.1 at 76.2%, GDPval-AA at 1656 Elo, MCP Atlas at 83.6%.

Google claims it delivers frontier-level intelligence at less than half the cost of comparable frontier models. That's a direct competitive shot at OpenAI and Anthropic pricing.

Gemini 3.5 Pro is already in internal use at Google and is expected to roll out next month, according to blog.google.

Gemini Spark: The Quiet Big Deal

MindStudio's breakdown of I/O 2026 correctly identifies Gemini Spark as the most underreported announcement of the conference.

Spark is a small model purpose-built for on-device inference — no cloud round-trip required. It runs natively on Android and ChromeOS. Google is positioning it as the default inference layer for Pixel devices going forward, according to MindStudio.

The benchmark that matters: sub-50ms first-token latency on mid-range Android hardware. That's fast enough to feel instant in a UI. For developers building voice or real-time apps, that number is significant.

Spark also introduces "Adaptive Context," which lets the model scale down gracefully on lower-end hardware instead of crashing out. That's good engineering, not just good marketing.

The practical implication: AI assistants that work offline, medical tools where patient data can't leave the device, industrial edge applications. An entire category of apps that weren't viable before now are, according to MindStudio.

What Mainstream Coverage Is Missing

The tech press — left-leaning outlets included — has fixated on the singularity rhetoric and the Omni brand name.

What deserves more attention:

First, Spark's on-device capabilities represent a genuine architectural shift away from cloud dependency. That has implications for privacy, cost, and latency that matter to regular people — not just developers.

Second, the SynthID watermarking built into Omni is a real accountability mechanism in the deepfake era. That deserves serious coverage, not a footnote.

Third, Gemini 3.5 Flash's cost claim — frontier performance at under half the price of competitors — represents a direct challenge to OpenAI's business model. If it holds up under independent benchmarking, that's a major market story.

Fourth, Google confirmed Gemini 3.5 Pro is already running internally and ships next month. The AI model competition isn't slowing down.

What Happened at I/O 2026

Google didn't just talk at I/O 2026. They shipped Omni Flash, benchmarked 3.5 Flash against the competition, and quietly laid the groundwork for an on-device AI ecosystem with Spark.

The singularity framing was theater. The product releases were real.

Whether the benchmarks hold up outside Google's own slides — and whether Omni's physics simulation actually improves past stuffed animals flipping upside down mid-skydive — the coming months will test those claims.