Gemini Spark Hits Google AI Ultra in the US: How It Works

Google’s Gemini Spark, the company’s 24/7 personal AI agent, started reaching United States subscribers of its Google AI Ultra plan this week, moving from a trusted-tester preview into open beta. Spark runs on Google’s cloud around the clock, automating multi-step jobs across Gmail, Calendar, Docs and a growing list of outside apps, and it keeps going even when your phone is switched off.

Most of the early coverage fixes on the chores Spark can take off your plate: booking a table, drafting an email, planning a trip. Two numbers say more about what Google actually built, a hard ceiling on how many jobs it will run at once and a 60% cut to the subscription that unlocks it.

What Gemini Spark Does and Who Can Use It

Spark is a background worker, not a chatbot. You hand it a goal, it plans the steps, then it carries them out on Google-run servers and checks back with you before anything sensitive happens. On the web, a new “Spark” tab sits opposite “Chat” in the side panel. On Android and iOS it lands between Search chats and Daily brief, marked “Beta.”

• Beta, US only: English-language, ages 18 and up, on the web and the Gemini mobile apps.
• Cloud-resident: it keeps working when your devices are off, running on Gemini 3.5 Flash and Google Antigravity.
• Bundled with Ultra: Google’s AI Ultra plan, repriced at I/O to start at $100 a month, down from $250.
• Hard cap: up to 15 tasks running at any one time.

Google began seeding Spark to trusted testers the week of May 19 and opened broader US beta access the week of May 26. You can read the feature set on Gemini Spark’s official overview page, which still labels wider availability as coming soon.

Tasks, Schedules and Skills: The Three Parts

Everything Spark does breaks into three pieces, and learning the vocabulary is the difference between a useful agent and a confused one. You define the task (what you want), set the schedule (when it should run), and supply the skill (how to do it).

Tasks: The Goal You Hand Over

A task is the high-level objective. “Plan and manage my business trip to London” is a task, as is “give me a custom news digest on the stories I follow.” You describe it in plain language in a “Describe your task” field that also accepts file uploads and a notebook. A work panel then tracks the planned, current and completed steps so you can watch Spark think.

You can also step in. A take-over control lets you grab the remote browser mid-job, finish a step yourself, and hand it back.

Schedules: When Spark Wakes Up

A schedule decides when a task runs without you asking. It can be a fixed time (“every day at 8 AM, give me an update on AI news”) or a condition (“when my flight is delayed, propose an updated itinerary”). These are separate from the scheduled actions inside the regular Gemini chat experience, and they are what make Spark feel persistent rather than reactive.

Skills: Reusable Instructions

A skill is a saved “set of reusable instructions and additional context” that teaches Spark how to do something and which tools to use. You call one with @ or /, or let Spark pick it automatically when it fits. The often-cited example: ask Spark to read your last 50 outgoing emails, build a personal writing-style guide, and apply it every time it drafts a message. In the London trip, a “Travel Booking” skill and a “Gmail Writing” skill can fire together to rebook a room and send the confirmation.

The Apps Spark Can Reach Into

Spark’s reach is what separates it from a smarter autocomplete. Inside Google Workspace it can manage your calendar, search and summarize Gmail threads, draft and forward replies, build and format Sheets with formulas, generate Slides from a prompt, and create or edit Docs. Connections are off by default, so you choose what it touches.

Beyond Google, Spark reaches outside apps through the Model Context Protocol (MCP, an open standard that lets AI agents talk to third-party services in a consistent way), which the Model Context Protocol specification from Anthropic introduced in late 2024. A remote browser with automatically saved sign-in info lets it click through sites and, for example, add items to a cart, while a remote computer handles code execution.

So a single prompt can pull notes from a Doc into a Canva deck, read context from your inbox to make an OpenTable reservation, or assemble an Instacart order, the kind of app-hopping that used to mean typing the same details three times.

Why the 15-Task Ceiling Matters

Here is where the marketing and the machinery part ways. Spark shares the same compute-based usage limits as the rest of Gemini, where quota refreshes every five hours until you hit a weekly cap, and exhausting a larger model quietly drops you to a smaller one. On top of that sits the agent-specific limit that tells you the most.

• You can have up to 15 tasks running at a given time.
• You have to wait for tasks to finish before making another request.
• A schedule will not fire if 15 tasks are already running, so an overloaded queue can silently skip your 8 AM news digest.

Persistent agents burn the budget that matters here, which is compute, not storage. A model that wakes on a schedule, drives a remote browser, runs code and checks in for approval is doing real work on Google’s servers whether or not you are watching. The ceiling is the tell: always-on autonomy is expensive enough that even a $100-a-month tier gets a queue limit. The flip side is supervision, which Google’s own documentation puts plainly.

Your supervision is the most important way to protect against risk.

That sentence sits in Google’s Gemini Spark support documentation, alongside confirmation prompts for sensitive actions and a list of prohibited tasks. An agent that can reach your inbox, your files and your saved logins is powerful in exactly the way that makes data handling the open question, the same tension running through Apple’s auto-delete privacy setting tied to its Gemini deal.

How Spark Fits Google’s Agent Push

The price move is the strategy in plain sight. At I/O, Google reset AI Ultra to start at $100 a month, down from $250, and trimmed its top tier to $200, a 60% entry-price cut that lands the same week Spark opens to US beta. Cheaper seats only pay off if people use the agent enough to stay, which is why Spark is the headline draw of the cheaper plan. The $100 tier also stacks 20 terabytes of cloud storage, YouTube Premium and a higher Gemini usage limit on top.

Google is not building this in a vacuum. The Model Context Protocol it adopted is the same open standard rivals are coalescing around, and the summer roster of Adobe, Samsung, Spotify, GitHub, Notion and Slack signals a land grab for whichever agent becomes the one users route their day through. Owning Gmail, Calendar and Drive gives Google a starting position competitors cannot copy.

That advantage is widening. Gemini’s footprint now stretches well past Google’s own apps, including a deal that puts the model inside a redesigned Siri, as covered in Gemini’s expansion across 2.5 billion Apple devices. An agent that already lives where you work, and a model that already lives on your phone, is a hard combination to dislodge.

For now the proof is narrow: US, English, 18-plus, and a 15-job queue. Spark is a real agent doing real work, with a real bill attached.

If the beta convinces enough Ultra subscribers to let an agent run their inbox and calendar unattended, the price cut pays for itself and the summer app partners arrive into a habit already formed. If the queue limits and approval prompts make it feel more like supervised software than a true assistant, Google will be selling autonomy that users keep choosing to do themselves.

Frequently Asked Questions

Is Gemini Spark free to use?

No. Spark requires a Google AI Ultra subscription and is currently limited to subscribers in the United States who are 18 or older. Google reset the Ultra plan at I/O to start at $100 a month, down from a previous $250.

What model does Gemini Spark run on?

Spark runs on Gemini 3.5 Flash paired with Google Antigravity, executing on Google’s cloud servers rather than on your device, which is why it can keep working when your phone or laptop is off.

How many tasks can Gemini Spark run at once?

Up to 15 at a time. You must wait for running tasks to complete before submitting another, and a scheduled action will not start if 15 tasks are already in progress.

Can Gemini Spark work with apps outside Google?

Yes. Through the Model Context Protocol, Spark connects to Canva, OpenTable and Instacart at launch, with Adobe, Samsung, Spotify, GitHub, Notion and Slack expected to follow over summer 2026.

Where do I find Gemini Spark in the app?

On the web, a “Spark” tab appears opposite “Chat” in the side panel. On Android and iOS, Spark sits between Search chats and Daily brief and is marked “Beta.”

Is Gemini Spark available outside the United States?

Not yet. The beta is US-only and English-only at launch, available on the web and the Gemini mobile apps, with no announced date for wider international or language support.