Google Gemini Spark Launches 24/7 AI Agent in 2026

Google just handed its most loyal subscribers a coworker that never sleeps. At I/O 2026, the company unveiled Gemini Spark, a 24/7 agentic assistant you talk to through its own Gmail address, and the beta is now reaching AI Ultra subscribers in the United States. The detail that should make every rival nervous is not the agent itself. It is that Google quietly halved the price of the plan it runs on.

What Actually Happened

Gemini Spark is a personal agent built on Gemini 3.5 and powered by Antigravity, Google's expanded harness for building agentic software. Alphabet CEO Sundar Pichai framed it as the next evolution of the digital assistant, a system designed to take on long-horizon tasks with minimal oversight rather than answering one prompt at a time. Unlike a chatbot that lives inside a tab, Spark runs on dedicated virtual machines on Google Cloud, which means it can keep working in the background even after you close your laptop. The pitch is simple and unusually concrete: you hand it a multi-step job, walk away, and come back to a finished result.

The interface is the part that breaks convention. Spark gets its own dedicated Gmail address, and you assign it work the same way you would email a human colleague. It pulls context automatically from Gmail, Google Docs, Sheets, and Slides without manual setup, browses the open web through Chrome, and can draft replies on your behalf. Crucially, Spark also supports the Model Context Protocol, the open standard for connecting agents to external tools, so it is not boxed inside Google's own suite. An enterprise version, Gemini Spark in Gemini Enterprise, extends the same background worker across Workspace, custom connectors, and third-party systems.

Then came the price. Google cut its top AI Ultra tier from $250 to $100 per month, the plan Spark ships on first. AI Ultra now bundles five times the usage limits of the AI Pro plan, 20 terabytes of cloud storage, and YouTube Premium. The beta is rolling out to US Ultra subscribers now, with broader availability to follow. In a single keynote, Google reframed what a premium AI subscription buys you: not a smarter chatbot, but a tireless agent plus a 60 percent price drop on the bundle that contains it.

Why This Matters More Than People Think

The headline reads like a product launch, but the real move is a repositioning of where AI value lives. For two years the industry sold access to intelligence by the prompt and by the token. Spark sells outcomes measured in completed tasks. When the unit of value shifts from "answer my question" to "finish my project," the entire pricing logic of the category changes, and Google is the first hyperscaler to put that bet in front of consumers at scale rather than burying it in an enterprise pilot.

Distribution is the second reason this lands harder than a spec sheet suggests. Google does not have to convince anyone to install Spark. It reaches billions of people through Gmail, Chrome, Android, and Workspace, the exact surfaces Spark already plugs into. A startup shipping an identical agent would spend years and hundreds of millions acquiring the users Google can reach with a settings toggle. The email-as-interface trick is deceptively powerful here, because every knowledge worker already lives in an inbox and already knows how to delegate by writing one.

The price cut reframes the competitive math for everyone else. OpenAI's ChatGPT Pro sits at $200 per month and Anthropic sells Claude Max at comparable tiers. Google just undercut both while adding an agent, 20TB of storage, and a YouTube Premium subscription that alone retails near $14 a month. For a household already paying Google for storage and video, the marginal cost of a frontier agent is now close to zero. That is not a feature war. It is a deliberate attempt to make the competition's flagship subscription look overpriced overnight.

There is a structural reason Google can absorb this price where others cannot. Alphabet runs Spark on its own TPU fleet and its own data centers, so the compute that powers a long-running agent costs Google a fraction of what a rival pays renting Nvidia capacity through a cloud markup. Anthropic and OpenAI buy most of their inference at someone else's margin, which means a $100 agent subscription that is comfortable for Google could be ruinous for them. Vertical integration, from the TPU silicon up through Gmail at the top, is the quiet weapon here. The keynote sold an assistant, but the balance sheet sold a moat: Google is the only company that owns every layer the agent touches, and that is what lets it treat a frontier subscription as a loss leader for data rather than a profit center it cannot afford to discount.

The Competitive Landscape

OpenAI is the obvious target. Its Operator and ChatGPT agent features pioneered the browser-driving assistant, and its Dreaming V3 memory upgrade, which began reaching Plus and Pro users on June 4, pushes in the same long-horizon direction. But OpenAI lacks the native inbox, the browser, and the document graph that make Spark feel less like a demo and more like a colleague. Anthropic, fresh off a $65 billion round at a $965 billion valuation and a confidential IPO filing, has the strongest underlying model in Claude Opus 4.8, yet it has no consumer distribution surface of its own and still reaches most users through partners and APIs.

Microsoft is the closest structural rival because it owns the other great email and document franchise. Copilot already lives inside Outlook and Microsoft 365, and the company is racing to reduce its dependence on OpenAI with in-house models. The difference is posture: Microsoft sells Copilot mostly to enterprises through per-seat licensing, while Google is leading with a consumer agent and a slashed consumer price. Whoever wins the household first may set the default that enterprises later inherit, the way personal Gmail seeded corporate Workspace adoption a decade ago.

The historical parallel is Google Assistant in 2016, which arrived with enormous reach and underwhelming autonomy and never became the agent Google promised. The lesson cuts both ways. Reach without reliability produced a decade of "set a timer" jokes, and Spark inherits the same trust deficit. But the underlying models are categorically more capable now than the intent-matching systems of 2016, and the failure of Assistant taught Google exactly which overpromises to avoid. The company that fumbled the first assistant era has the most to prove and the most context about how it went wrong.

Hidden Insight: The Inbox Is the New Operating System

The most underappreciated decision in Spark is that Google made email the control plane for an autonomous agent. This is not a cosmetic UI choice. Email is asynchronous by nature, which matches how long-horizon agents actually work: you fire off a request, the agent grinds for minutes or hours on a cloud VM, and the result lands when it is ready. A chat window trains users to expect an instant reply and punishes any task that takes real time. An inbox trains users to expect a considered response later. Google picked the interface that fits the technology instead of forcing the technology into a chat box that fights it.

There is a deeper strategic reason to route an agent through Gmail. Every email Spark sends and receives becomes structured, addressable context that compounds over time. The agent's work product lives in the same thread-and-label system that already organizes a user's life, which means Spark accumulates memory and accountability for free. You can audit what it did by reading the thread. You can forward its work to a human. You can reply to redirect it. None of that requires Google to build a new memory architecture, because the inbox has been a durable, searchable memory layer for twenty years.

This also quietly solves the agent trust problem better than a dashboard ever could. The deepest barrier to autonomous agents is not capability, it is the fear of handing a system the keys and not knowing what it will do. Email is the most familiar accountability surface most people own. When the agent's actions arrive as messages you can read, question, and reverse, the psychological cost of delegation drops sharply. Google is betting that trust in agents will be won through legibility, not through better benchmarks, and the inbox is the most legible interface knowledge workers have. The bear case, however, is straightforward: an agent that can read your inbox, drive your browser, and act for hours unattended is also the richest target prompt injection attackers have ever been handed, and a single poisoned email or web page could turn Spark against the user it serves. Critics argue Google is racing distribution ahead of safety, and the security model is the part the keynote said least about.

The price cut, read through this lens, is not generosity. It is customer acquisition for a data flywheel. Every Spark user generates a stream of real-world task data: what people actually delegate, where the agent fails, which corrections humans make. That feedback is the rarest training fuel in the industry, far more valuable than another web scrape, and Google is effectively paying users to produce it by making the subscription cheaper than the storage and video it already includes. The $100 price is the cost of buying the world's largest agentic task dataset, and it may be the best money Google spends this year.

Consider what that dataset actually captures. A web crawl tells a model how humans write about doing things. A stream of delegated tasks tells a model how humans actually want things done, including the messy back-and-forth of clarifying a vague request, the corrections when an agent overreaches, and the unspoken preferences that only surface when real work is on the line. This is the gap between knowing the answer and knowing the job, and no amount of pretraining on public text closes it. By putting a free agent in front of millions of motivated users, Google is generating the one corpus its competitors cannot scrape, license, or buy, because it does not exist anywhere until a user delegates a real task and reacts to the result. The price cut is the acquisition cost for that corpus, and the corpus is what compounds into a lead no benchmark can show.

What to Watch Next

In the next 30 days, watch the reliability reports from beta users in the United States. The single number that matters is task completion rate on multi-step jobs, because a 24/7 agent that fails silently is worse than no agent at all. Watch also whether OpenAI or Anthropic responds to the $100 price with cuts of their own. If they hold their $200 tiers, Google's bundle pressure compounds. If they match, the consumer AI subscription has effectively halved in price in a single quarter, a deflation that will reshape every revenue model in the sector.

Over 90 days, the leading indicators are enterprise pickup of Gemini Spark in Gemini Enterprise and the breadth of MCP connectors third parties build. An agent is only as useful as the systems it can touch, and the size of the connector ecosystem will reveal whether developers treat Spark as a platform or a curiosity. Watch security disclosures closely too. An agent with inbox access, browser control, and a standing cloud VM is an enormous attack surface, and the first credential exfiltration or prompt injection incident involving Spark will set the regulatory tone for the whole agentic category.

By 180 days, the verdict question is retention. Novelty drives the first month of any AI launch. The real test is whether users keep delegating real work after the demo wears off, and whether Google can show that Spark handled tasks people would otherwise have paid an assistant or a freelancer to do. If retention holds, expect Google to push Spark into Android as a system-level agent and to start charging enterprises premium rates for the same capability it gave consumers at a discount. Watch for the moment the price stops falling and starts climbing again, because that will mark the point Google believes the habit is locked in.

Google did not just launch an agent. It cut the price of the future in half and dared everyone else to follow.

---