On May 14, an APK leak named a feature nobody had heard of: Gemini Spark. Five days later, Sundar Pichai walked on stage at Google I/O 2026 and confirmed it — Spark is real, it ships next week to U.S. AI Ultra subscribers, and the leaked privacy line (“may share your info or make purchases without asking”) got softened but didn’t disappear.
This is what Spark actually is now that the keynote is over: a cloud-resident personal AI agent that runs 24/7 on a dedicated Google VM, plugs into Gmail/Docs/Slides plus 30-plus third-party services via MCP, and lives on the new $100/month AI Ultra tier. It can book restaurants. It can put in an Instacart order. It will draft your inbox replies while you sleep. It cannot, contrary to the leaked beta screen, spend money without asking — at least not anymore.
Here’s the practical version of what Spark is, what it can do today, what’s still aspirational, and whether the $100/month is worth it for you specifically.
What just launched, in plain language
At I/O 2026 on May 19, Google announced four things at once: Gemini 3.5 Flash (the new default model), Gemini Omni (multimodal generation), Antigravity 2.0 (the agent runtime), and Gemini Spark (the user-facing agent that runs on top of all three). Spark is the consumer-friendly product. Antigravity is what’s actually executing under the hood.
The mental model: Antigravity is the engine. Spark is the car you drive. You tell Spark in plain English what you want done — “watch my inbox for anything from the legal team and summarize it once a day” — and Antigravity spins up the right sub-agents, calls the right MCP connectors, and runs the task on a Google Cloud VM that doesn’t need your laptop awake.
Three things distinguish Spark from regular Gemini in the app:
- It runs in the cloud, not on your device. Your phone can be locked, your laptop closed, the Wi-Fi at home off — Spark keeps going. This is the headline shift.
- It’s persistent. A regular Gemini chat ends when you close the tab. Spark holds onto multi-day tasks (“track these five flights for the next two weeks and ping me if any fare drops 15%”) and resumes them every time it wakes up.
- It can act, not just answer. With your permission, it books, orders, drafts, sends, schedules, and follows up. The “answer the question” surface of Gemini becomes “do the work” with Spark.
That’s the leap. And that’s why the $100/month price tag (more on that in a moment) is positioned as the floor — Google wants Spark out of the experimental ghetto and into the hands of anyone willing to pay.
How Spark works under the hood
Spark runs on the new Gemini 3.5 Flash model wrapped in Antigravity 2.0, Google’s “agent harness.” Think of Antigravity as the operating system for agents: it does the orchestration, the tool calls, the sandboxing, the credential management, and the long-running task scheduling. Spark is one application built on top of it. Google also ships Antigravity as a desktop app, a CLI, and an SDK so developers can build their own agents on the same plumbing.
The integrations work through MCP — Model Context Protocol, the open standard Anthropic introduced last year that everyone in the agent space has now adopted. Each connected app (Canva, OpenTable, Instacart, plus Workspace) is exposed to Spark as an MCP server. Spark calls the server, gets back structured tool definitions, and uses those to take action. No raw credentials are passed to the language model itself — that’s done by the MCP runtime in a separate sandbox. This is the architectural reason Google is willing to let an agent act on your behalf at all.
Google said “30+” third-party apps will be available at or near launch. The keynote and follow-up coverage named these specifically:
| Connector category | Apps named at launch |
|---|---|
| Workspace (first-party) | Gmail, Docs, Slides, Calendar |
| Design | Canva |
| Restaurants | OpenTable |
| Shopping & groceries | Instacart |
| Coming this summer | Local file access via macOS desktop app, custom sub-agents you can train, text/email Spark directly |
The macOS desktop client is the bridge that lets Spark touch your local files — it indexes a folder you authorize, then makes those files available to the cloud agent through the same MCP protocol. Same security model as the third-party connectors, just running on your machine instead of someone else’s server.
How much it costs and who gets it
Spark is included with Google AI Ultra. At I/O, Google also restructured the Ultra plan — the old single $249.99 tier is gone, replaced with two:
| Plan | Price | What you get | Spark included? |
|---|---|---|---|
| Google AI Pro | $19.99/mo | Gemini app with Gemini 3.5, 2 TB storage, Workspace integrations | ❌ |
| AI Ultra (new entry) | $100/mo | Everything in Pro + 5× usage limits, priority Antigravity access, Gemini Spark, 20 TB storage, YouTube Premium | ✅ |
| AI Ultra (top tier) | $200/mo | Everything above + 30 TB storage, deepest model access, highest agent concurrency | ✅ |
The $100/month tier is the new mid-band — explicitly positioned for “developers, creators, and power users.” That’s the floor of entry for Spark. Project Mariner used to be the Ultra-only experimental feature; Mariner is now retired and its browser-control work has been folded into Spark’s playbook.
Geographic availability at launch is the U.S. only. Google has not committed to an EU or U.K. date, and the EU AI Act timing makes that a deliberate hesitation rather than an oversight — the obligations for general-purpose AI models with agentic capabilities are still being clarified at the European AI Office. Expect the EU rollout to be measured in months, not weeks.
How Spark spends your money (and the protocol behind it)
This is the part that got everyone’s attention from the leak. The original beta onboarding said Spark “may do things like share your info or make purchases without asking.” That language is gone from the production product. What replaced it is the Agent Payments Protocol — AP2 — which Google is also rolling into Google Shopping later this year.
AP2 is a permission framework with three controls you can’t bypass:
- Per-transaction approval. Spark cannot complete a purchase without your explicit confirmation. The confirmation can be a tap in the Gemini app, a Siri-style voice “yes,” or a desktop notification — but it has to happen.
- Spending caps. You set a per-transaction limit (e.g., $50), a daily limit, and a category-by-category allowlist (groceries: yes; subscriptions: ask; everything else: never).
- Merchant restriction. You can scope Spark to specific merchants (Instacart yes, random Shopify storefront no). The default is conservative.
Google’s VP of Labs, Josh Woodward, described it in briefings as “giving an agent your wallet is like giving a teenager their first debit card — you set the controls before you hand it over, not after.” The analogy works. The protocol also creates a durable audit trail for every transaction, so if a dispute happens you have a complete record of what Spark did, when, and with whose approval.
This is a meaningful change from the leaked behavior. If you read the May 14 onboarding screen and noped out, the May 19 production version is materially different — purchases require approval, and there’s a public protocol you can audit.
What this means for you
The same product reads very differently depending on who you are. Quick rundown:
If you’re a heavy Google Workspace user (Gmail + Docs + Calendar): Spark is going to be the most useful agent of the three majors for you, because the deepest integrations are with the tools you already live in. The $100/month is a real number, but if you do an hour of inbox triage a day or a lot of meeting prep, the math gets honest fast.
If you’re already paying for Claude Pro or ChatGPT Plus: Spark adds a different shape of help, not a redundant one. Claude Cowork ($20/mo via Pro) and ChatGPT Operator are both more conversational; Spark is more “set it and forget it.” The right question isn’t which to use — it’s whether you want both a hands-on AI assistant and a background AI worker, and at what total cost.
If you’re a solo operator (freelancer, consultant, one-person business): This is the audience Spark was built for. Background workflows — inbox triage, weekly client digests, invoice chasing, RSVP tracking, content scheduling — are exactly what the $100/month buys you. It’s not a luxury for this group; it’s an FTE replacement for the operational sliver of the work.
If you’re in the EU or U.K.: Don’t expect Spark on day one. Google’s hesitation is real and the AI Act compliance review is non-trivial for an autonomous agent. If you’re in Berlin or London, watch for a Q3 2026 update at the earliest.
If you handle sensitive client data (lawyers, accountants, therapists, doctors): Read the AP2 controls and the data-access list before you opt in. The default behavior is more conservative than the leak suggested, but Spark still reads from Gmail and Calendar by default. If your inbox contains privileged communications or PHI, set Spark’s data sources to only the accounts and folders you’ve sanitized, and turn off the Personal Intelligence profile aggregation. The controls exist; they’re just off by default for the wrong reasons.
If you tried Project Mariner and gave up: Spark is the second-generation version of what Mariner promised. Mariner’s biggest limitation was that it lived in Chrome and required you to leave the browser open; Spark moves all of that into Google’s cloud VMs. The same web-control tasks that frustrated you in Mariner — the unreliable form-filling, the timeouts, the “lost session” errors — should be substantially smoother now that the agent has its own dedicated runtime.
What Spark still can’t do
Even with the launch behind us, the agent space has gaps. Spark is not yet a finished product:
- It can’t operate without an internet connection. The whole point is the cloud VM. Offline work goes through the regular Gemini app, not Spark.
- It can’t move money. The Plaid-style read-only model that ChatGPT just shipped for finance? Spark doesn’t have an equivalent for banking yet. It can manage your calendar, your shopping, and your communication; it cannot pay your bills.
- It can’t act inside apps that don’t have an MCP connector. Your favorite niche scheduling tool, your industry-specific CRM, the law firm’s billing system — if there’s no MCP server, Spark can’t touch it. Expect the connector library to grow, but slowly.
- It can’t be customized with your own sub-agents yet. Google said “this summer” for custom skills you train yourself. At launch (next week), you’re using the workflows Google ships.
- It can’t run in the EU, U.K., Canada, Brazil, India, or Japan at launch. U.S.-only beta, period.
- It can’t replace the Pro plan. If you wanted Gemini for chat-style help on the cheap, you still need AI Pro at $19.99. Spark is in addition to that, not instead of.
The bottom line
Gemini Spark is now real, it ships next week to U.S. AI Ultra subscribers, and it’s a substantially different product than the leaked beta suggested. The autonomous-purchase concern was the right thing to flag in May — and it’s also the part Google actually changed before launch. If you read the leak coverage and bailed, the production version deserves a second look.
The honest take: the always-on agent category is going to ship from all three majors this quarter. Spark, Claude Cowork, and ChatGPT Operator are converging on the same product shape. Spark’s bet is that being inside Google’s app stack — and being cheaper than the old Ultra plan — wins enough mainstream users to define the category. We’ll see whether the $100/month feels like a fair trade six months from now.
If you want to get fluent in this stack before everyone at your job assumes you already are, two of our courses map directly to where Spark fits:
- Gemini Personal Intelligence at Work — real workflows for inbox triage, calendar prep, and the cross-app stuff Spark is built to handle. The right course if you’re saying yes to the $100/month and want to make it pay for itself.
- Gemini Personal Intelligence Privacy Playbook — the controls, the data sources, the AP2 protocol in plain English. The right course if you read this post and still aren’t sure about handing an agent your Gmail.
Sources
- Google blog: The Gemini app becomes more agentic
- Google blog: Everything new in our Google AI subscriptions
- TechCrunch: Google introduces Gemini Spark, a 24/7 agentic assistant
- Tom’s Guide: Google’s new $100 AI Ultra plan changed the AI race
- Engadget: The Google AI Ultra plan now starts at $100 a month
- Android Authority: Google drops its most expensive AI Ultra pricing
- CNBC: Google unveils AI model Gemini 3.5 and AI agent Gemini Spark
- Decrypt: Google Launches Gemini Spark — A 24/7 AI Agent
- 9to5Google: Everything Google announced at I/O 2026
- Gizmodo: Google Comes for OpenClaw With Gemini Spark
- TestingCatalog: Google unveils 24/7 Gemini Spark AI Agent
- Mashable coverage: Agent Payments Protocol and Spark spending controls
- EU AI Act Service Desk: Implementation timeline
