Item: Gemini Spark
Author: Elena Marchetti

Google has spent the last two years playing catch-up in the AI assistant market. Gemini Spark, unveiled at I/O 2026 on May 19, is the company's clearest response - not a faster model or a cheaper API, but a different product category: an AI agent that runs 24 hours a day without needing your laptop open.

TL;DR

7.2/10 - A real architectural step forward for personal AI agents, undercut by early-beta limits and unsettling fine print
Cloud-persistent execution is a genuine innovation: Spark works while you sleep, not just while you chat
US-only beta, opaque usage caps, and a buried disclaimer saying it "may make purchases without asking" are real concerns
Best for: heavy Google Workspace users comfortable delegating to an AI. Skip if you value data minimalism or live outside the US.

The pitch is simple. Instead of an assistant that answers questions when you ask, Spark takes on long-horizon tasks - parsing your inbox for overdue replies, populating a Sheets tracker from calendar invites, booking a restaurant through OpenTable - and does it whether you're at your computer or not. That's not a small distinction. Every personal AI assistant before Spark, including OpenAI's Workspace Agents, has operated within the scope of a single session. Spark breaks that assumption.

Whether it breaks anything else is the more interesting question.

What Spark Actually Is

Spark runs on Gemini 3.5 - the same model family launched last week - paired with a custom agentic harness called Google Antigravity. Antigravity handles the orchestration layer, routing tasks across Workspace APIs and MCP-connected third-party services, while the Gemini backbone handles the reasoning. The whole system runs on dedicated virtual machines in Google Cloud, not on your device.

That cloud-native architecture is the product's defining technical decision. It means Spark keeps working after you lock your phone or step away from your desk for a week. It also means Google holds a detailed execution log of everything the agent did on your behalf - every email it drafted, every site it browsed, every purchase it authorized.

May 19, 2026 - Gemini Spark announced at Google I/O 2026 with a Gemini app redesign called Neural Expressive.
Week of May 26, 2026 - Beta rolls out to US Google AI Ultra subscribers at the $100/month tier.
Summer 2026 - macOS desktop app integration planned; broader MCP integrations and custom sub-agent creation expected.

Core Capabilities

Spark's native integrations cover the full Google Workspace stack: Gmail, Docs, Sheets, Slides, and Calendar. The announced use cases include parsing credit card statements for hidden subscription charges, synthesizing scattered meeting notes into polished documents, and building block-party logistics across calendar invites and a Sheets tracker. These are the demos Google showed at I/O - worth noting, because Spark isn't yet available for independent testing at scale.

Beyond Workspace, Spark connects to third-party services through MCP (Model Context Protocol) - the open standard that Google and the rest of the industry adopted earlier this year. Launch partners are Canva, OpenTable, and Instacart, with Google saying "30+" integrations will be available at or near general release. The connections are API-driven, not screen-scraping: Spark calls MCP servers, gets back structured tool definitions, and acts through those. That makes it more reliable than computer-use approaches, but also more constrained - if a service hasn't built a MCP server, Spark can't touch it.

On mobile, you can send Spark a task via a dedicated Gmail address. Android Halo - a new status overlay on Android - shows live progress for background work.

The redesigned Gemini app with Neural Expressive design language and Spark integration The Gemini app's Neural Expressive redesign, launched alongside Spark at Google I/O 2026. Source: blog.google

The Architecture Gap vs Competitors

Stacked against the two closest alternatives, the competitive picture looks like this:

Feature	Gemini Spark	Claude Cowork	ChatGPT Agent
Always-on execution	Yes (cloud, 24/7)	No (desktop session)	No (browser session)
Architecture	API-driven MCP	Screen automation	Browser extension
Native integrations	Google Workspace	Any visible app	1,000+ GPT plugins
Third-party scope	30+ at launch	Anything on screen	Very broad
Availability	US beta only	Generally available	Generally available
Pricing entry point	$100/month	Varies	$20/month

Spark's 24/7 cloud execution is a real architectural advantage over session-bound competitors. Claude Cowork can operate across more arbitrary apps because it reads the screen directly - but it stops the moment you close the window. ChatGPT Agent has the broadest plugin ecosystem but runs in a browser tab. Spark is the only option that keeps going when no one is watching.

That advantage narrows much once you consider what "30 integrations" means in practice. Most professional workflows touch software that won't have a MCP server at Spark's launch, let alone in six months. A legal team using a niche contract management tool, a startup on a custom CRM, an academic using specialized research software - none of these get Spark's autonomous reach. They get a very capable inbox assistant.

The Privacy Problem

Google's I/O keynote highlighted Spark's safeguards: explicit opt-in, per-app permission controls, confirmation required before sending emails or spending money. All of that's accurate.

What the keynote skipped is in the onboarding fine print. The actual product disclaimer reads that Spark "may do things like share your info or make purchases without asking." That clause sits in direct tension with the marketing language about approval requirements, and Google hasn't explained the conditions under which Spark bypasses its confirmation flow.

A persistent agent with access to your email, calendar, and payment authorization is a very high-value attack target. Prompt injection through a malicious email isn't theoretical.

The data exposure goes deeper than the disclaimer. Enabling Spark gives a single persistent agent access to your full email history, documents, calendar, and browser activity. Google logs every task Spark executes. That's a qualitatively different privacy posture than asking a chatbot a question - the exposure is continuous, not transactional.

Google says Spark includes adversarial prompt detection to guard against injection attacks, but hasn't published a security architecture document or third-party audit. A single malicious email that tricks Spark into forwarding attachments or confirming a purchase would be a serious incident. For a broader look at what makes AI agents risky, this is a textbook case.

A professional workspace environment with multiple screens - the kind of workflow Spark targets Spark is designed for knowledge workers managing email, documents, and scheduling across interconnected services. Source: pexels.com

Usage Caps: The Undisclosed Ceiling

Spark's availability isn't binary. Google moved the Gemini app from daily prompt counts to a compute-used model that factors in prompt complexity, features used, and conversation length. Heavy Spark use draws down your allowance faster than chatting. When you hit the limit, Spark stops working. The reset schedule isn't publicly disclosed, and there's no option to purchase additional compute.

That structure makes Spark unreliable for any workflow you'd call business-critical. You might schedule inbox monitoring every morning, then discover mid-Wednesday that Spark is offline until an opaque reset resets. For $100 to $200 a month, that's a meaningful gap between the marketing and the product.

Pricing and Availability

Spark is available on two Google AI Ultra tiers, both US-only at launch:

$100/month (new tier introduced at I/O) - 5x usage limits vs Pro, 20 TB storage, YouTube Premium, Spark access
$200/month (existing flagship tier) - same Spark access, higher overall limits

Both tiers are US-only at beta launch. European availability will likely trail by months - the EU AI Act's transparency obligations for consumer-facing AI agents take effect in August 2026, and Google will need time to adapt disclosures for that regulatory context.

Strengths

Cloud-persistent execution is a genuine step beyond session-bound agents
Deep Workspace integration through native APIs, not screen-scraping
MCP architecture is extensible and broadly adopted across the industry
Android Halo gives visible task progress without opening the app
Confirmation flow for high-stakes actions is the correct default

Weaknesses

US-only beta severely limits who can test and report on it honestly
Onboarding disclaimer contradicts the "always confirm" marketing language
Opaque usage caps with no way to purchase additional compute
Only 30 third-party integrations at launch - most professional tools aren't covered
No published security architecture or independent audit of prompt-injection defenses
Google holds a continuous, detailed log of all agent activity

Verdict

Gemini Spark is the most architecturally coherent personal AI agent Google has shipped. The cloud-execution model is a real innovation, and the Workspace integration is deep where it works. If you live in Google's ecosystem and you're willing to trust Google with persistent, continuous access to your accounts, this is the agent to try.

The trust ask is major. The product is early. The usage cap structure makes Spark unreliable for critical workflows. The privacy fine print is real and the marketing glosses over it. And the US-only beta means most of the world is assessing Spark from announcement slides and demo videos rather than first-hand experience.

Google has built something worth paying attention to. Version 1.0, though, is still version 1.0.

Score: 7.2/10

Gemini Spark Review: Google's Always-On AI Agent