Don't Count Gemini Out Yet

Everyone said Google lost I/O.

I watched it. I read the post-mortems. I scrolled the X discourse, half “Google’s behind on raw benchmarks,” half “where’s the OpenAI killer?” Went in with the same priors most people had. Google’s playing catch-up. Gemini’s a generation behind. The model wars are over and Sam won.

Then I actually started looking at what they shipped.

And what I saw wasn’t a company chasing. It was a company consolidating.

For the last eighteen months, Google has had the same problem every big-tech AI lab has: too many products, too many overlapping tools, too many user-facing brands, all chasing different parts of the same stack. Gemini CLI here. Gemini in Workspace there. Maps doing its own AI thing. Search bolting on AI Overviews. Assistant slowly getting absorbed. Android Studio gaining agentic features. Each ship independently, each fighting for the same engineering and product attention, each with subtly different mental models.

At I/O 2026, that ended. Google took the disparate AI tooling ecosystem (the CLIs, IDEs, models, voice products, the mobile apps, search, the OS, and a maps service two billion people use every month) and folded them into one coherent product story.

That’s not retreat. That’s a moat being dug in real time.

Five things I saw at I/O that nobody should be sleeping on.

1. Gemini CLI is now Antigravity CLI (and SDK)

Let’s start with the dev tools, because that’s where the consolidation is most visible.

Gemini CLI is sunset. June 18, 2026, less than a month from now as I write this, it stops serving Google AI Pro and Ultra users. The thing that replaces it, Antigravity CLI, isn’t just a rename. It’s a ground-up rewrite with a different architecture and a different theory of what a terminal coding agent is for.

The technical changes:

Rewritten in Go. The old Gemini CLI was Node-based: slower startup, heavier memory. Antigravity CLI is snappy from cold. Matters when you’re using it dozens of times a day.
Async subagents are first-class. You can fire a long-horizon refactor task or a multi-source research task at a background subagent and keep working in the main session. The CLI doesn’t block. Multi-agent orchestration isn’t a hack pattern anymore. It’s the default mental model.
Shared engine with Antigravity 2.0 desktop. Sessions are portable between the terminal and the GUI desktop app. Start something in the terminal at your desk, hand it off to the GUI on another machine, come back to the terminal. The session is the unit, not the surface.
First-class extensibility. Agent Skills (reusable Markdown slash commands), Hooks (lifecycle interceptors), MCP servers (both stdio and HTTP), and Plugins (the rebranded Gemini CLI Extensions) all survive the migration. Existing investment isn’t lost.
Multi-model out of the box. Default is Gemini 3.5 Flash. Depending on your plan, you can also run Gemini 3.1 Pro, Claude Sonnet, Claude Opus, GPT-OSS 120B. The CLI is no longer a Gemini-only experience.
Command mode. agy -p "your prompt" --output-format json. Non-interactive. Pipeable. Scriptable. Runs in cron, runs in CI. This is the part most people are missing: Antigravity CLI isn’t just an interactive coding agent for sessions at your desk. It’s a runtime component you can wire into pipelines.

And there’s the Antigravity SDK alongside it, which lets you build on the same agent harness without using the CLI at all. Skills, hooks, subagents, MCP: programmatically, in your own code, against your own infrastructure.

The pricing detail matters. Antigravity CLI is included on Google AI Pro. Many developers already pay for AI Pro for Gemini access. Now they get a Claude Code-class agentic CLI in that subscription. Compare that to xAI’s pricing model: Grok Build requires SuperGrok Heavy at $300/month, or $99/month for the first six months as an introductory rate.

Bundling is a strategy. Google is playing it deliberately.

2. Real-time voice is the default modality now, across the entire product line

This is the one I think the AI Twitter discourse keeps under-weighting, and I think it’s actually the most strategically important shift at I/O.

Voice isn’t a feature Google launched. Voice is a default modality Google is wiring into everything.

Gemini Live is the obvious one. Google updated it with wave-form UI for visual feedback, added privacy and compliance UX flags, full-duplex audio that doesn’t feel like turn-taking. But Live isn’t the story by itself. The story is the defaulting:

Voice-driven navigation in Maps (which I’ll get to in a second).
Voice in the Gemini app, including walking and biking contexts where typing is literally impossible.
Voice in Workspace as an editing modality: not just dictation, but actual edits to your slides and docs in conversational form.
Voice in Android, where the Assistant lineage is being absorbed into Gemini as a single coherent agent across every Google surface.

Why this matters: the products that read well as voice are different from the products that read well as text and taps. Search-as-conversation. Hands-busy contexts. Navigation. Long-running personal tasks. Background workflows. Older users who don’t want to type.

Those are categories where the device that handles voice naturally wins. And Google has the products designed around voice already: Search, Maps, Assistant, Android. OpenAI has ChatGPT. Anthropic has Claude. Both excellent, both fundamentally screen-based.

If voice becomes the default way most people use AI in three years (which I think it will), Google’s distribution advantage compounds. They have the surfaces. The competitors don’t.

3. Google Maps got a voice, and you can actually talk to her

If you haven’t tried Ask Maps, do it. Rolled out March 12, 2026 in the US and India, on Android and iOS.

You ask Maps real-world questions in real-world language:

“My phone is dying, where can I charge it without waiting in a long line for coffee?”

“Is there a public tennis court with lights I can play at tonight?”

And it answers, using your prior searches, your saved places, your patterns of use. The answer is tailored to you. Not generic. Not a list of nearest matches. An answer.

Navigation got the same upgrade. Voice guidance is more natural: “Go past this exit and take the next one for Illinois 43 South.” Landmarks instead of just street names: “Turn right after the Thai Siam Restaurant.” The immersive 3D view actually looks like what’s in front of you, with road details and natural voice guidance.

It also works while you’re walking or biking now, which it didn’t before. That rolled out earlier this year. You can talk to Maps the same way you’d talk to a friend who’s lived in the neighborhood for ten years.

This is what consolidation looks like in practice. Maps isn’t a separate AI product Google bolted Gemini onto. It’s an interface to Gemini that happens to know where you are, where you’ve been, and what you tend to want. The model is upstream. The product surface is downstream. The user experiences the surface, not the model.

That’s the right architecture. And it’s the architecture every major Google product is moving toward.

4. Gemini Spark: your 24/7 personal agent on a Google VM

This is the headline. The thing that, once it’s actually rolling, I think most consumer AI is going to be measured against.

Spark is Google’s swing at the personal agent. The one that just runs, all the time, without you opening a laptop.

What it actually is:

Runs on a dedicated Google Cloud VM under your account. 24 hours a day, 7 days a week. Your laptop stays closed. Your phone is in your pocket. Spark keeps working.
Built on Gemini 3.5 and the Antigravity agent harness: the same underlying agent infrastructure as the developer tools, but consumer-packaged.
Plugged into Gmail, Docs, Sheets, Slides, plus 30+ third-party services via MCP: Calendar, Instacart, OpenTable, and so on.
Can book a restaurant. Place an Instacart order. Draft your inbox replies overnight while you sleep. Pull facts from scattered emails and docs and write a quote response for you. Manage your calendar.
Cannot move money. No banking integration. Probably the right call for v1. The failure modes of an autonomous agent that can transfer money are catastrophic in a way that the failure modes of an agent that books dinner reservations are not.
$100/month on the Google AI Ultra plan. Trusted testers now; AI Ultra subscribers in the US next week.
Not available in the EU, UK, Canada, Brazil, India, or Japan at launch. Regulatory friction is real, and Google is being conservative about it.

The pricing matters. $100/month is positioned firmly above the consumer AI subscription tier (ChatGPT Plus at $20, Gemini Advanced at $20) and below the developer/power-user tier (Antigravity for devs is bundled in AI Pro at $20; SuperGrok Heavy is $300). It’s targeting people who would pay for an executive assistant.

How Spark stacks up against OpenClaw and Hermes Agent

The interesting competitive question isn’t Spark vs ChatGPT or Spark vs Claude. It’s Spark vs the open-source personal-agent stack (OpenClaw and Hermes) that took 2026 by storm.

Quick refresher on those:

OpenClaw is the broad-integration framework. 374,000 GitHub stars as of this writing. Self-hosted. Open-source. Dev-first. Skills are static files you write, version, and maintain yourself. You’re in full control, but you’re also doing the work.

Hermes Agent is Nous Research’s project. 163,000 GitHub stars, but it’s the fastest-growing agent framework of 2026, going from 8.8k stars to 163k in three months. Its differentiator: self-generating skills. You don’t write skills manually. The agent creates them autonomously during use, refines them, and compounds them across sessions. Their tagline, “the agent that grows with you,” is accurate. Internal benchmarks from Nous in April 2026 showed self-generated-skill agents completing complex research and code-execution tasks 40% faster than fresh non-learning instances. As of May 2026, Hermes has overtaken OpenClaw on OpenRouter’s global daily rankings: 224 billion tokens per day to OpenClaw’s 186 billion.

Spark is neither of these. Spark is the consumer-managed agent:

You don’t host it. (Google does, on a dedicated VM.)
You don’t write its skills. (Google’s agent harness handles that.)
You don’t tune it. (You set preferences; Google handles the rest.)
You hand Google your Gmail and Google handles everything.

That’s a fundamentally different bet. OpenClaw and Hermes are agents for builders. Spark is an agent for everyone else. The people who will never run a CLI but will absolutely pay $100/month to have something competent draft their inbox while they sleep.

And that audience is two orders of magnitude bigger than the OSS developer audience. The Hermes folks and the OpenClaw folks should not feel threatened by Spark. The product lines barely overlap. But the AI news cycle is going to treat Spark as if it’s competing head-on with Hermes for the same user, and that misread is going to let Google build a consumer moat in plain sight while everyone is busy comparing GitHub star counts.

The thing to watch isn’t whether Spark beats Hermes on some benchmark. It’s how many of the 300 million people who already pay Google something get up-sold to Ultra in the next twelve months.

5. New models (Gemini 3.5 Flash, Gemini Omni) vs Nemotron and Qwen

Gemini 3.5 Flash launched at I/O on May 19. The pitch: frontier intelligence at flash-tier speed and cost. The benchmarks Google highlighted are unusually specific:

Terminal-Bench 2.1: 76.2% for agentic coding tasks executed in a real terminal.
GDPval-AA: 1656 Elo on Google’s economically-valuable-tasks benchmark.
MCP Atlas: 83.6% on Model Context Protocol tool use.

Notice what’s being measured. Not GPQA Diamond. Not MMLU. Not raw reasoning. Agentic execution. Tool use. Terminal work. The benchmarks Google chose to highlight are the benchmarks that map to what developers actually use models for in 2026.

And the number that should make you sit up: 3.5 Flash beats Gemini 3.1 Pro on those benchmarks. The new Flash model beats the previous Pro model on the things developers actually use models for. That is not the same shape of progress as chasing GPQA points.

Gemini Omni is the other release. The pitch: “create anything from any input, starting with video.” A leap in world understanding, native multimodality, and conversational editing. Google’s demo was video-to-video: generating an 8-second clip from a text prompt, then editing it conversationally (“add more energy at second four,” “make the lighting golden hour”). If you’re a creator, this is a different category of tool than what’s been shipping.

The honest comparison

Now let’s be honest about where Gemini doesn’t win.

Qwen 3.5-Omni (Alibaba, open-weights) has 215 SOTA results across audio, audio-video understanding, reasoning, and interaction benchmarks. The Qwen 3.5-Omni Plus variant outperforms Gemini 3.1 Pro on general audio understanding, reasoning, and translation tasks. And, critically, Qwen is open-weights. You can run it on your own hardware. You can fine-tune it. You can deploy it without a Google account or an API key.

Nvidia Nemotron 3 Nano Omni is a specialized multimodal efficiency model. It completes an iterative 5-round video tagging workflow in 8.30 hours, notably faster than comparable models. Nemotron doesn’t lead on raw reasoning benchmarks. It’s not trying to. It’s a workhorse for specific multimodal pipelines at scale, the kind of thing that lives in enterprise video processing systems.

Does Gemini win every benchmark? No. Qwen out-points on audio. Nemotron out-runs on video-pipeline efficiency. If your shop is benchmark-shopping per axis, you’re not picking Gemini for everything.

But here’s the kicker. Qwen is open-weights, fantastic for builders, but Alibaba is not landing Qwen-Omni on 2 billion Android phones. Nemotron is enterprise infrastructure, invaluable for media companies, but Nvidia is not waking Nemotron up at 3 AM to draft your inbox replies as Spark.

The model is not the product. The distribution is the product. Google still owns the distribution.

What else mattered

A few things that didn’t make the five-point outline but matter for the bigger picture:

Antigravity 2.0 desktop app + cloud sessions. You can start a session in the cloud, hand it off to your terminal, and back. The session is the unit, not the surface. This is the right abstraction for agentic work that takes hours and crosses devices.
MCP everywhere. Antigravity speaks it natively. Spark speaks it. The CLI speaks it. The open-protocol bet Anthropic made with MCP is paying off for Google because it lets them plug into the same ecosystem builders are already wiring up.
Gemini in Workspace as a real editing partner, not a sidebar. Generative slides. Sheet transformations. Doc rewrites, all in-line.
Search has Gemini inline by default now, not just as AI Overviews. Full agentic search for complex queries that previously took five tabs.

The take

The Google-is-behind narrative is the easy take. The harder take, and I think the right one, is that Google spent the last eighteen months consolidating disparate AI products into a coherent agentic stack, and at I/O they shipped the receipts.

They didn’t win every benchmark. They don’t need to. They’re playing a different game.

In Google’s game:

The model is plumbing, not the product.
The platform is the moat.
The distribution is already paid for.

The CLIs and SDKs are unified under Antigravity. The models are competitive on the benchmarks that matter for real work. The personal agent is real and shipping next week. The voice-first surfaces are everywhere Google’s users already are. The MCP integration story is genuine. The pricing strategy bundles power tools into existing subscriptions.

I’m not going to tell you to short OpenAI or short Anthropic. I’m a Claude user. I’m building on the Anthropic stack right now. They’re going to be fine.

But I am going to tell you: don’t count Gemini out yet.

The next twelve months belong to whoever can make agentic AI a default consumer feature, not a developer toy. Right now Google is the only player with the distribution to do that at scale.

I’ll be running Spark the day it lands in the US. If you’re building agents (open-source or not, developer-facing or consumer-facing) you should be paying attention to what Google is teaching its users to expect.

Because once those expectations are set, that’s the bar.

Catalyst for this post: this I/O recap on YouTube. Worth the watch.

Sources: Google I/O 2026 recap · Gemini CLI → Antigravity CLI transition · Antigravity 2.0 platform · Gemini Spark (TechCrunch) · Ask Maps (TechCrunch) · Hermes Agent vs OpenClaw (MarkTechPost) · Qwen 3.5-Omni review · Nemotron 3 Nano Omni review