Gemini Spark: Google's $100/Mo 24/7 AI Agent (2026)
Gemini Spark is Google's always-on AI agent, launched May 19, 2026 in Google AI Ultra ($99.99/mo). It manages Gmail, Calendar, Docs, and bookings 24/7.
Gemini Spark is a semi-autonomous AI agent from Google that runs 24/7 on dedicated cloud servers, managing your Gmail, Calendar, Docs, and Drive even when your devices are off. Bundled into Google AI Ultra at $99.99 to $200 per month (down from $249.99), it drafts emails for your approval and connects to Canva, OpenTable, and Instacart, though it's US-only and English-only during its May 2026 beta.
Gemini Spark is Google's 24/7 AI agent, launched May 19, 2026 at I/O, running persistently on dedicated Google Cloud virtual machines so it keeps working even when your phone is off. It's bundled into Google AI Ultra, now $99.99 per month after a 60% price cut from $249.99. Spark manages Gmail, Calendar, Docs, and Drive, and drafts replies for your approval, currently in US-only English beta.
Maker: Google · Autonomy: semi autonomous · Maturity: BETA
Underlying models: Gemini 3.5 Flash
About Gemini Spark
Gemini Spark is a 24/7 AI agent from Google, announced on May 19, 2026 at Google I/O alongside the Gemini 3.5 Flash model. Unlike the standard Gemini chat assistant, which stops working the moment you close the tab, Spark runs persistently on dedicated Google Cloud virtual machines with a lifecycle measured in hours or days. It is designed to solve a specific problem: keeping up with email, scheduling, and document work without you having to open an app and ask for help every time. Spark sits inside the Gemini app as the proactive, always-on counterpart to Google's reactive chat assistant, and it launched as part of a broader shift Google described as moving Gemini from a tool that surfaces data to an agent that acts on it. Spark's reasoning runs on Gemini 3.5 Flash, which Google says processes around 284.2 tokens per second, about 4x faster than comparable frontier models. That speed is wrapped in Antigravity 2.0, an internal Google agent runtime that adds goal persistence, task decomposition, tool orchestration, safety constraints, and state recovery on top of raw model calls. This is what makes Spark an agent rather than a chatbot: it can hold a goal like keeping an inbox triaged across days, decide when to act, call tools like Gmail and Calendar, and recover its state if interrupted. Consequential actions, such as sending an email, are held as drafts for human approval, making Spark semi-autonomous rather than fully autonomous. Spark is best suited to people who already live inside Google Workspace and want help with high-volume, repetitive digital admin. A small business owner can let Spark monitor incoming customer emails in Gmail's AI Inbox, generate personalized draft replies using context pulled from past emails, Docs, and the calendar, and surface a Daily Brief each morning summarizing what needs attention. It can create calendar events directly from email content, summarize and generate documents in Docs, Sheets, and Slides, and at launch it connects via the Model Context Protocol (MCP) to Canva for design work, OpenTable for restaurant bookings, and Instacart for grocery orders. Executive assistants and solo founders are the clearest fit; enterprise IT teams that need narrowly scoped, audited agent permissions are a weaker fit, since security researchers have flagged Spark's broad, always-on Workspace access as carrying execution risk on top of normal data exposure risk. Spark is not sold separately. It is bundled into Google AI Ultra, and Google used Spark's launch to cut Ultra's price by 60%, from $249.99 to $99.99 per month, specifically to give the new agent a mass-market price point. A higher AI Ultra Premium tier costs $200 per month (down from $250) with roughly 20x the usage limits of the $20/month AI Pro plan, versus about 5x for the $99.99 tier; both Ultra tiers also include 20TB of storage and YouTube Premium. As of its May 2026 launch, Spark is available only to US-based, English-language Google AI Ultra subscribers, rolling out first to trusted testers and then more broadly to Ultra subscribers. There is no public developer API for Spark itself, and no free tier. Google has said more MCP-based integrations are coming over the summer of 2026, including GitHub, Notion, Slack, Adobe, Samsung, Spotify, and CapCut, on top of the Canva, OpenTable, and Instacart connectors available at launch. Broader rollout to Google AI Pro subscribers and Workspace business previews is also planned for later in 2026, though international availability outside the US will be shaped in part by the EU AI Act's consumer-agent obligations, which take effect on August 2, 2026. Given Google's pace of shipping Gemini app updates, expect Spark's connector list and rollout footprint to expand quickly through the rest of 2026, though as of the beta launch it remains a US-only, Workspace-centric agent rather than a general-purpose, model-agnostic one.
Pricing
Not sold separately. Bundled into Google AI Ultra: $99.99/mo entry tier (cut 60% from $249.99 at Spark's launch, ~5x AI Pro usage limits, 20TB storage, YouTube Premium) or AI Ultra Premium at $200/mo (down from $250, ~20x AI Pro limits). No free tier and no standalone Spark pricing.
Key Features
- 24/7 Persistent Runtime: Spark runs on dedicated Google Cloud virtual machines with a lifecycle measured in hours or days, continuing tasks even when your phone is locked or laptop is shut, unlike session-based chat agents.
- Gmail AI Inbox: Spark monitors incoming email and writes personalized draft replies using context from your Gmail, Docs, and Calendar, holding each draft for your approval before anything sends.
- Daily Brief: A continuously updated digest pulled from Gmail, Calendar, and Tasks surfaces what needs your attention each morning without you having to ask.
- Cross-App Workspace Reasoning: Spark cross-references Gmail, Calendar, Docs, Sheets, Slides, Drive, and Maps, for example creating calendar events directly from email content.
- MCP Third-Party Connectors: At launch Spark connects to Canva, OpenTable, and Instacart over the Model Context Protocol, with GitHub, Notion, Slack, Adobe, Samsung, Spotify, and CapCut planned for summer 2026.
- Built on Gemini 3.5 Flash: Spark's reasoning runs on Gemini 3.5 Flash at roughly 284.2 tokens per second, about 4x faster than comparable frontier models, wrapped in the Antigravity 2.0 agent runtime.
Strengths
- Runs persistently on dedicated Google Cloud VMs for hours or days, so tasks keep progressing even when your phone is off or laptop is shut, unlike tab-bound agents.
- Bundled into Google AI Ultra after Google cut the plan's price 60% from $249.99 to $99.99 per month, lowering the cost of an always-on agent.
- Native, no-setup access to the full Google Workspace stack (Gmail, Calendar, Docs, Sheets, Slides, Drive, Maps), so it can act on real account data from day one.
Weaknesses
- Available only as a US-based, English-only beta since its May 19, 2026 launch, with no international rollout date announced.
- Requires a Google AI Ultra subscription starting at $99.99/month; there is no free tier, standalone price, or public API for Spark itself.
- Spark's always-on Workspace access raises permission risk: security researchers note it can inherit broad folder and group access in Gmail, Docs, and Drive without evaluating business context.
Frequently Asked Questions
What is Gemini Spark and what does it do?
Gemini Spark is a 24/7 AI agent from Google, announced at Google I/O on May 19, 2026, that runs persistently on dedicated Google Cloud virtual machines rather than ending when you close a chat tab. It is built on Gemini 3.5 Flash, which Google says runs at 284.2 tokens per second, roughly 4x faster than comparable frontier models, wrapped in an agent runtime called Antigravity 2.0. Spark monitors your Gmail, Google Calendar, Docs, Sheets, Slides, and Drive and takes action on your behalf, such as drafting email replies, creating calendar events from email content, and generating documents. It also produces a Daily Brief, a continuously updated digest of what needs your attention pulled from Gmail, Calendar, and Tasks. At launch it connects to three third-party services over the Model Context Protocol (MCP): Canva, OpenTable, and Instacart. Spark is currently a US-only, English-only beta available to Google AI Ultra subscribers. Unlike a standard chatbot that waits for a prompt, Spark continues working across hours or days even when your phone is locked or your laptop is off.
How much does Gemini Spark cost in 2026?
Gemini Spark is not sold as a standalone product; it is bundled into Google's AI Ultra subscription. At I/O 2026, Google cut the price of AI Ultra by 60%, from $249.99 to $99.99 per month, specifically to give Spark a mass-market price point. A second tier, AI Ultra Premium, costs $200 per month (down from $250) and offers up to 20x the usage limits of the $20/month AI Pro plan, compared to roughly 5x for the $99.99 tier. Both Ultra tiers include 20TB of cloud storage and a YouTube Premium subscription alongside Spark access. There is no free tier for Spark, and the lower-cost AI Pro ($20/month) and AI Plus plans do not include it at launch. Google has not announced a separate, lower-priced Spark-only plan. Because Spark requires an existing Google account to act on, business users may also need a Workspace subscription on top of AI Ultra.
Is Gemini Spark fully autonomous?
No, Gemini Spark is semi-autonomous: it monitors your accounts and plans actions continuously, but consequential steps require your approval. For example, when Spark drafts a reply in Gmail's AI Inbox based on context from your emails, Docs, and Calendar, it holds that draft for you to review before anything is sent. Background monitoring tasks, like the Daily Brief that scans Gmail, Calendar, and Tasks, run without prompting and do not require approval. Spark's persistence is what sets its autonomy apart: it runs on dedicated Google Cloud virtual machines with a lifecycle measured in hours or days, maintaining goal state across that time in a way a single API call cannot. Security researchers have flagged that this always-on access to Gmail, Docs, Sheets, and Drive carries execution risk on top of normal exposure risk, especially where permission scopes are broad or inherited. Compared with fully autonomous coding agents, Spark sits closer to a proactive personal assistant that still keeps a human in the loop for anything that leaves your account.
What AI model powers Gemini Spark?
Gemini Spark runs on Gemini 3.5 Flash, the model Google announced alongside Spark at I/O 2026. Google states Gemini 3.5 Flash processes around 284.2 tokens per second, about 4x faster than comparable frontier models, which suits Spark's need to run many background tasks cheaply and continuously. Gemini 3.5 Flash is wrapped in Antigravity 2.0, an internal agent runtime that adds goal persistence, task decomposition, tool orchestration, safety constraints, and state recovery on top of the raw model calls. Users cannot choose a different underlying model for Spark; it is a fixed, managed configuration rather than a model-agnostic agent framework. There is no published fine-tuning detail specific to Spark beyond the Antigravity wrapper. Because Gemini 3.5 Flash and other 2026 Gemini models support roughly 1 million token context windows, Spark can hold large amounts of Workspace context, such as emails, documents, and calendar history, when reasoning about a task. Google has not disclosed a separate model card specifically for the Spark configuration as of its May 2026 beta launch.
What are the best alternatives to Gemini Spark?
Claude Code from Anthropic is the closest comparison on hokai for autonomous execution, though it is built for software engineering rather than personal Workspace tasks; pick Claude Code if your priority is an agent that writes, tests, and ships code rather than managing email and calendar. ChatGPT Agent from OpenAI combines a virtual-computer Operator mode with deep research and works across web, mobile, and desktop; choose it if you want broader platform coverage and don't need Spark's Google-native Workspace depth. Manus AI is a fully autonomous browser-and-code agent that some 2026 comparisons rate ahead of ChatGPT Agent on output quality; pick Manus if you want an agent that can browse, code, and manage files without Google account integration. Anthropic's Claude-based Cowork-style assistants are reported around $20 per month, far below Spark's $99.99 entry price, making them a choice for freelancers on a tighter budget who still want business-tool depth. In short, Spark wins on persistent Google-native execution, Claude-based tools win on business-tool depth at lower cost, and ChatGPT wins on platform breadth, and many teams in 2026 are running more than one of these in parallel.
Who is Gemini Spark best for?
Gemini Spark is best for people already living inside Google Workspace, such as solo founders, small business owners, or executive assistants who handle a high volume of email, scheduling, and document work. A typical use case is a small business owner who lets Spark monitor incoming customer emails, draft replies using context from past conversations and the calendar, and hold those drafts for a quick approval each morning alongside a Daily Brief. It also suits early adopters in the US who are comfortable testing English-only beta software and don't mind Google AI Ultra's $99.99 to $200 per month price. It is not a good fit for users outside the US, since Spark has no international rollout date and is English-only at launch. It is also not ideal for enterprise IT teams that need SOC2-audited, narrowly scoped agent permissions, since Spark's broad always-on Workspace access has been flagged as a permission risk by security researchers. Developers looking for a coding-focused agent with a public API should look elsewhere, since Spark has no standalone developer API at launch.
How does Gemini Spark compare on benchmarks?
Google has not published SWE-bench, WebArena, or GAIA scores for Gemini Spark itself, since it is a product wrapper (Antigravity 2.0) around the Gemini 3.5 Flash model rather than a benchmarked research agent. The one concrete performance figure Google has shared is for the underlying model: Gemini 3.5 Flash runs at roughly 284.2 tokens per second, which Google describes as about 4x faster than comparable frontier models, supporting Spark's need to run many background tasks economically. For agent-level comparisons, third-party 2026 testing has focused on rivals like Manus AI, GenSpark, and ChatGPT Agent on tasks such as web browsing and code execution, areas where Spark does not directly compete since it is scoped to Google Workspace and a handful of MCP-connected services (Canva, OpenTable, Instacart) at launch. Because Spark is still in a US-only beta as of May 2026, independent benchmark coverage is limited and likely to expand once it reaches Workspace business previews and a wider audience later in the year. Until then, the honest assessment is that Spark should be evaluated on task completion within Gmail, Calendar, and Drive rather than against general agent benchmarks.
How do you get started with Gemini Spark?
First, you need a Google AI Ultra subscription, which costs $99.99 per month for the entry tier (or $200 per month for AI Ultra Premium), and you need to be located in the US since Spark is English-only and US-only during its beta. Once subscribed, Google rolls out Spark access in stages, starting with trusted testers and then expanding to Ultra subscribers, so access may take up to a week or more after signing up. You reach Spark through the Gemini app, where you connect your Google account so it can read Gmail, Calendar, Docs, Sheets, Slides, and Drive. A good first task is to enable the Daily Brief, which starts generating a digest from your Gmail, Calendar, and Tasks without further setup. From there, try asking Spark to monitor your inbox and draft replies to a specific type of email; it will hold drafts for your approval in Gmail's AI Inbox rather than sending automatically. If you want to test third-party actions, Spark's launch MCP connectors for Canva, OpenTable, and Instacart let you try generating a design, booking a table, or ordering groceries through the same interface. No coding or API setup is required for any of this, since Spark is a consumer feature inside the Gemini app, not a developer product.