Gemini Spark Goes Live, AGI Hype Peaks, and Google's $1B Cost Claim Needs Scrutiny — The I/O Aftermath

What Actually Launched — Not Just What Was Announced

Gemini 3.5 Flash is live. As of Tuesday, May 19, it's the default model for the Gemini app and AI Mode in Search worldwide, according to Google. API pricing is $1.50 per 1M input tokens and $9 per 1M output tokens — cheaper than Gemini 3.1 Pro's $2/$12 entry price, per Ars Technica.

Gemini Spark is rolling to testers this week. Beta access for Google AI Ultra subscribers ($100/month tier) starts next week in the U.S., according to TechCrunch. It is NOT broadly available yet. Most coverage buried that detail.

Gemini Omni Flash is live today for all Google AI Plus, Pro, and Ultra subscribers, per Engadget. The fuller Omni model and 3.5 Pro are still coming — Pro reportedly next month.

Project Genie just went global for adult AI Ultra subscribers worldwide, following its January U.S. debut. The new Street View integration launches for U.S. locations first, per Engadget.

The $1 Billion Claim — Read the Fine Print

Sundar Pichai told reporters that companies running roughly one trillion tokens per day on Google Cloud could save more than $1 billion annually by shifting 80 percent of workloads to Gemini 3.5 Flash, according to VentureBeat.

One trillion tokens per day. That's not your average enterprise. That's a hyperscaler-level customer. The claim is technically accurate for a vanishingly small number of organizations and is being reported broadly as if it applies to the market generally. It doesn't.

What IS broadly applicable: 3.5 Flash generates output at 4x the speed of comparable frontier models, per Engadget, and benchmarks from Artificial Analysis cited by VentureBeat show it outperforming Gemini 3.1 Pro — Google's own flagship from roughly five months ago — on nearly every major test. Those numbers are real. The $1 billion headline is marketing.

Hassabis Said What?

DeepMind CEO Demis Hassabis closed the keynote by declaring: "When we look back at this time, I think we will realize that we were standing in the foothills of the singularity."

The Verge correctly flagged the contradiction. Just months ago, Hassabis told Bloomberg "we're nowhere near" the singularity by his own definition — and his definition of singularity differs from the classical Vinge/Kurzweil version anyway. He's defining it essentially as full AGI. His stated timeline: 50 percent chance of AGI by 2030.

This got ONE sentence of stage time at the keynote. The tech press largely let it pass. Wired pushed back — Hassabis told them he thinks firing engineers over AI productivity gains is "a lack of imagination." His exact quote: "If engineers are becoming three or four times more productive, then we just want to do three or four times more stuff."

That's a reasonable position. It's also one that directly contradicts what Amazon, Salesforce, and Block have been saying publicly as they use AI as cover for layoffs.

The Privacy Trade That Nobody Is Pricing In

Gemini Spark's entire value proposition depends on reading your Gmail, Calendar, Docs, Slides, and Sheets — and eventually your browser and local files. The Verge noted that Google's advantage over OpenAI and Anthropic isn't the AI — it's that Google already has all your data behind a simple opt-in menu.

That opt-in framing is doing a lot of work. Google first integrated Gemini into Workspace in 2024. It launched "Personal Intelligence" in January — Gemini reasoning across Gmail, Photos, Search, and YouTube history without prompting, per The Verge. Millions are already using it, according to Josh Woodward, VP of Google Labs.

VentureBeat raised what seems obvious: what happens when Spark gets it wrong? The outlet flagged "urgent questions about trust, spending guardrails, and what happens when an artificial intelligence agent misinterprets a user's intent." CNBC noted Google didn't discuss privacy terms for the new smart glasses at all.

Conservatives worried about government surveillance have reason to care equally about a private corporation's always-on agent reading every email. The data doesn't belong to Google — it belongs to you.

The Subscription Restructure — Who Actually Benefits

Google restructured its AI plans at I/O. The breakdown, per Engadget and ZDNET:

• AI Plus: ~$20/month. Gets Gemini 3.5 Flash, Omni, basic features.
• AI Pro: ~$20/month upgraded tier. Adds YouTube Premium Lite, eventually Google Pics.
• AI Ultra (new lower tier): $100/month. 5x usage limits vs. Pro, Gemini Spark access, 20TB storage.
• AI Ultra (full): $200/month, down from $250. 20x usage limits, Project Genie access.

The $50 price cut on the top tier is real. So is the storage cut — from 30TB to 20TB at that level. Gemini Spark is ONLY for the $100+ tiers. Information agents in Search go first to Pro and Ultra subscribers this summer, per TechCrunch.

Translation: most of what was demoed on Tuesday costs at least $100/month to actually use.

What Mainstream Coverage Got Wrong

Left-leaning outlets covered the privacy implications more than right-leaning ones did, but even they undersold it. ZDNET and VentureBeat were the most useful on actual enterprise numbers.

Gemini 3.5 Pro is still internal only. The model Google says will act as the "orchestrator" for the whole agentic system doesn't exist in your hands yet. The full vision Pichai described — Flash as sub-agent, Pro as planner — is architecture, not product. Not yet.

Also buried: AI Mode has already passed one billion monthly users, per ZDNET. Google didn't emphasize that. They should have. It's the most important adoption number they have.

What's Real and What Isn't

Google shipped real products Tuesday. Gemini 3.5 Flash is live, it's fast, and the pricing is genuinely competitive. The smart glasses hardware is real — Wired and Engadget both handled it. Universal Cart and Gemini Spark are in early rollout.

But Hassabis declaring the singularity's foothills while his boss sells billion-dollar cost savings to customers who don't exist at that token volume is exactly the kind of hype that makes normal people distrust all of this.

Verify the benchmarks. Check the fine print on opt-ins. And remember: when the product is free — or framed as a convenience — you're the data source, not just the customer.