How is Gemini Spark different from the Gemini app?

The Gemini app is a chat-style assistant — you ask, it answers. Gemini Spark is autonomous and persistent: you give it a goal ('book me a flight, find the textbooks from my syllabus, monitor flight prices to Berlin'), and it works in the background, surfacing approval prompts when it needs to take real-world actions. Spark also integrates third-party services via MCP, where the Gemini app does not.

How does Gemini Spark compare to ChatGPT Atlas and Claude agents?

Three different shapes of agent. Gemini Spark is a cloud-resident always-on agent that orchestrates across apps via MCP and asks for per-action approval. ChatGPT Atlas is OpenAI's agentic browser — it operates as the user inside a browser session. Claude agents (via the Claude Agent SDK and Managed Agents) are developer-first, designed to be deployed as task workers. Spark fits 'personal assistant', Atlas fits 'browser pilot', Claude fits 'agent backend'.

How do I get Gemini Spark?

Gemini Spark requires the Google AI Ultra plan ($99.99/month tier or higher). It's rolling out in beta to US subscribers within the week of May 19, 2026. Other regions follow. Spark needs explicit per-action approval for any external action (sending email, making purchases, modifying calendars), and you can revoke MCP connections at any time.

Quick Answer

What Is Gemini Spark? Google's 24/7 Personal AI Agent (May 2026)

Q: What is Gemini Spark?

Gemini Spark is Google's 24/7 personal AI agent, announced at Google I/O 2026 on May 19, 2026. It's powered by Gemini 3.5 Flash and Google's Antigravity agent framework, and runs continuously in Google Cloud. Spark can plan and execute multi-step tasks across Gmail, Google Workspace, and 30+ third-party apps connected via the Model Context Protocol (MCP). Each action requires user approval. It launches in beta for Google AI Ultra subscribers in the US within a week of the announcement.

Published: May 20, 2026

What Is Gemini Spark? Google’s 24/7 Personal AI Agent (May 2026)

Gemini Spark is Google’s always-on personal AI agent, announced at Google I/O 2026 on May 19, 2026. It runs in Google Cloud, uses Gemini 3.5 Flash as its reasoning engine, and orchestrates tasks across Google products and 30+ third-party apps connected via MCP. Each action it takes requires your approval.

Last verified: May 20, 2026

Quick facts

Property	Value
Announced	May 19, 2026 (Google I/O 2026)
Vendor	Google
Underlying model	Gemini 3.5 Flash
Framework	Google Antigravity
Where it runs	Google Cloud (24/7, not on-device)
App connections	30+ third-party tools via MCP
Required plan	Google AI Ultra ($99.99/mo or higher)
Initial rollout	US, beta, within a week of May 19
Action model	User approval required per real-world action

What Spark actually does

Spark is the productized version of an idea Google has been hinting at for two years: an assistant that doesn’t wait for you to ask. You give it goals or persistent contexts (“watch for flight deals to Berlin under €200,” “every Monday morning summarize the week ahead”), and it works on them in the background.

Example flows Google showed at I/O:

Class prep agent — “Find the syllabus in my Gmail, identify the required textbooks, find them on Amazon and Bookshop, add them to a cart.” Spark plans the steps and asks for one approval at the purchase step.
Trip planner — “Watch flights to Berlin in June under €200 and notify me; pre-fill a calendar hold once a candidate flight appears.”
Meeting bot — “Pull this week’s customer calls, summarize action items, draft follow-up emails (don’t send).”
Vendor monitor — “Track competitor pricing on these 5 SKUs and alert me when anything moves more than 5%.”

How Spark differs from a chat assistant

	Gemini app (chat)	Gemini Spark (agent)
Trigger	User asks	Goal-based, autonomous
Persistence	Per session	Always-on (24/7)
Multi-step planning	Reactive	Proactive
Third-party apps	Limited	30+ via MCP
Real-world actions	Suggestive	Executable (with approval)
Runs on	Device + cloud	Google Cloud
Cost	Free / Pro / Ultra	Ultra only ($99.99+)

Architecture (what Google revealed)

Reasoning engine: Gemini 3.5 Flash — chosen for speed in agentic loops.
Orchestrator: Google Antigravity framework (the same agent harness behind Antigravity 2.0 IDE).
Tool integration: Model Context Protocol (MCP), the same open protocol Anthropic introduced in late 2024. This is significant: Google is betting on MCP as the cross-vendor tool layer.
Safety layer: explicit per-action user approval, with audit log, MCP connection revocation, and SynthID watermarking on any generated media.
Compute: runs in Google Cloud — not on-device, unlike Gemini Intelligence on Android.

Gemini Spark vs ChatGPT Atlas vs Claude agents

Three big-vendor takes on “the agent”:

	Gemini Spark	ChatGPT Atlas	Claude (Agent SDK + Managed Agents)
Shape	Cloud-resident personal agent	Agentic browser	Developer agent runtime
Where it runs	Google Cloud, 24/7	Local browser session	Anthropic-managed or self-hosted
UX	Spark dashboard + approvals	Browser UI with agent mode	API, SDK, dashboards
Best for	Personal/ongoing tasks	Web tasks (forms, research, shopping)	Building agents for products
Required plan	Google AI Ultra	ChatGPT Plus / Atlas access	Pay-per-use API + plans
MCP support	Yes (30+ tools at launch)	Yes	Yes (original spec)

Privacy, safety, and approvals

Spark’s design centers on explicit approval for real-world actions: every email send, calendar change, purchase, or API call that affects the outside world surfaces an approval card. You can:

Pause / resume agents anytime.
Revoke MCP connections per app.
See a full audit log of agent actions and decisions.
Disable specific tools per agent.

Google also reiterated that Spark conversations are not used to train models for Workspace customers and that personal-account training settings carry over.

Pricing

Plan	Price	Spark access
Free	$0	Not included
Google AI Pro	$19.99/mo	Not included
Google AI Ultra	$99.99/mo	Yes (beta)
Google AI Ultra (top)	$199.99/mo	Yes, with higher concurrency

The $99.99 tier is new — Google created it specifically to give Spark a clean entry price. The $200 tier (previously $250) gets you Project Genie expanded access and 20x usage limits.

Limits and open questions

Region: US-only at beta launch; EU / UK / APAC dates not given.
Action model: per-action approval is strict — it will slow down high-volume agents but reduce accident blast radius.
MCP set: 30+ tools at launch (Google, Microsoft, Notion, Linear, Asana, Slack, GitHub, and others). The full list ships with the rollout email.
Personal data: Spark needs broad cross-app context to be useful. The trade-off is real.

TL;DR

Spark is Google’s “the model is the operating system” play. A cloud-resident, 24/7 agent that uses Gemini 3.5 Flash and Antigravity to coordinate across your apps, asking permission before it acts. It’s the most ambitious personal-agent product anyone has shipped to date — but it’s also Ultra-tier ($99.99/mo) and US-only at launch. If you’re already paying for ChatGPT Pro, this is the obvious head-to-head test.