Google on Tuesday unveiled Gemini Spark, a private AI agent designed to work across the clock — drafting emails, assembling paperwork, monitoring inboxes, and ultimately making purchases — even when a person's laptop computer is closed and their telephone is locked.
The announcement, made at Google I/O 2026, is the corporate's most formidable try but to rework its AI assistant from a device that solutions questions into one which autonomously completes duties. It additionally arrives at a second of extraordinary competitors, as Microsoft, OpenAI, Anthropic, and Apple all race to construct AI techniques that don't merely converse however act — finishing multi-step workflows with lowering human supervision.
"We’re in that a part of the cycle the place individuals need to see actual worth within the merchandise they use on a day-to-day foundation," Sundar Pichai, CEO of Google and Alphabet, stated throughout a press briefing forward of the keynote deal with. With Spark, he argued, that worth comes from an agent that by no means stops working. It operates across the clock in Google's cloud, he stated, so "you don't have to maintain your laptop computer open to ensure it's operating."
The product arrives at an inflection level for the know-how business, as Google, Microsoft, OpenAI, Anthropic, and Apple all race to construct AI techniques that don't merely converse however do — finishing multi-step workflows with lowering human supervision. It additionally raises pressing questions on belief, spending guardrails, and what occurs when a man-made intelligence agent misinterprets a person's intent.
Spark will start rolling out this week to a small group of trusted testers, with a beta deliberate for Google AI Ultra subscribers in america subsequent week.
Contained in the cloud structure that lets Gemini Spark work when you sleep
In contrast to standard AI assistants that activate solely when prompted, Gemini Spark is architecturally totally different. It runs persistently on Google Cloud infrastructure, powered by the corporate's new Gemini 3.5 Flash model and what Google calls the Antigravity agent harness — the identical underlying system that powers the corporate's inside developer instruments.
In sensible phrases, this implies Spark can settle for a posh instruction — "e-mail my boss a standing replace pulling the most recent figures from our shared spreadsheet and the mission timeline in our Slides deck" — after which execute it throughout a number of Google purposes with out additional enter. The agent can pull context from emails, paperwork, and calendar entries, synthesize the data, and produce a completed output.
Josh Woodward, VP of Google Labs, Gemini App, and AI Studio, described the expertise in visceral phrases in the course of the briefing: "If you use it, it nearly feels such as you're tossing issues over your shoulder — Spark's catching them and will get the job achieved."
The cloud-based structure is a deliberate design selection. As a result of Spark operates on distant servers relatively than on a person's system, it may possibly proceed working by way of duties after a person walks away. A pupil might ask Spark to construct a research information that updates itself as new assignments arrive from a professor. A small enterprise proprietor might instruct it to observe their inbox and flag potential buyer inquiries. A dad or mum might delegate the logistics of a neighborhood block occasion — monitoring RSVPs, coordinating contributions, scouting venues. These will not be hypothetical situations. Woodward stated they replicate how early testers have really been utilizing the product.
Over the approaching months, Google plans to develop Spark's capabilities considerably. The corporate will roll out MCP (Model Context Protocol) connections to greater than 30 third-party companions, together with Canva, OpenTable, and Instacart. Customers may even have the ability to textual content and e-mail Spark straight, create customized sub-agents for specialised duties, and join Spark to Chrome for web-based actions. Later this 12 months, a brand new Android interface known as Android Halo will present stay, at-a-glance visibility into what Spark is engaged on, displayed on the high of a person's telephone display.
Google compares its AI spending safeguards to giving a young person their first debit card
For all its ambition, Spark confronts a elementary problem that has bedeviled each AI agent thus far: How do you belief an autonomous system to behave in your behalf — notably when cash is concerned?
Google is conscious about the priority. When requested in the course of the press briefing how Spark would keep away from making unauthorized purchases, Woodward reached for an analogy that was hanging in its candor. "On the staff, we expect a number of it’s like when you're giving a young person their first debit card — there's form of limits and form of constraints round it, and that's how we'll be designing Spark as we undergo the 12 months," he stated.
At launch, Spark won’t autonomously make purchases. Customers will likely be given express alternatives to evaluation and approve any transaction earlier than it goes by way of. However Google has constructed the infrastructure for a extra autonomous future. Vidhya Srinivasan, who leads Google's adverts and commerce groups, launched the Agent Payments Protocol, or AP2 — a system designed to let AI brokers make safe purchases inside user-defined boundaries.
The idea works like this: a person tells their agent the precise manufacturers, merchandise, and spending limits they're comfy with. If the standards are met, the agent can mechanically full a purchase order. AP2 creates what Google describes as a clear, verifiable hyperlink between the person, the service provider, and fee processors, utilizing privacy-preserving know-how and tamper-proof digital mandates to make sure the agent is appearing inside its authorization. AP2 additionally generates a everlasting digital paper path, in order that if a return is required, the person and the service provider are trying on the identical report. Google plans to carry AP2 to its merchandise within the coming months, beginning with Gemini Spark.
The system is underpinned by the Universal Commerce Protocol (UCP), an open-source customary Google introduced earlier this 12 months that provides brokers and commerce techniques a typical language throughout the complete buying journey. The UCP Tech Council now consists of Amazon, Meta, Microsoft, Salesforce, and Stripe — a exceptional coalition that underscores how critically the business takes the prospect of agent-driven commerce.
Google additionally introduced the Universal Cart, an clever buying cart that works throughout retailers and Google providers. Customers can add gadgets whereas looking Search, chatting with Gemini, watching YouTube, or studying Gmail. The cart then works within the background — monitoring value drops, surfacing offers primarily based on fee card perks, and even flagging product incompatibilities. The buying infrastructure is rolling out within the U.S. this summer time throughout Search and the Gemini app, with YouTube and Gmail to observe.
How Google, OpenAI, Microsoft, Anthropic, and Apple are racing to construct the definitive AI agent
The announcement lands in the midst of probably the most intense aggressive interval in AI historical past. Google, Microsoft, OpenAI, Anthropic, and Apple are all racing to ship autonomous brokers that may do actual work — and every is putting a basically totally different architectural wager on how one can get there.
OpenAI just lately unified its Operator and deep research capabilities into ChatGPT agent — a system that brings collectively web site interplay, info synthesis, and conversational intelligence. It carries out duties utilizing its personal digital pc, shifting between reasoning and motion to deal with advanced workflows. The corporate emphasizes that customers stay in management, with ChatGPT requesting permission earlier than taking consequential actions. However the product has confronted scrutiny over reliability. OpenAI's Pc-Utilizing Agent scores 38.1% on OSWorld, the business benchmark for pc use duties, whereas people rating over 72%.
Anthropic launched its Claude Computer Use Agent in analysis preview in March, giving Claude the power to see, navigate, and management a person's desktop — clicking buttons, opening purposes, filling spreadsheets, and finishing multi-step workflows. Claude Cowork handles duties autonomously — customers give it a aim and Claude works on their pc, native information, and purposes to return a completed deliverable. Anthropic has iterated aggressively, just lately transport ten pre-built monetary brokers and pursuing deep Microsoft 365 integration.
Microsoft launched Copilot Cowork to maneuver past chat and into execution — serving to customers delegate actual duties and have them accomplished. Cowork runs within the cloud, that means customers don't have to fret about closing their laptop computer. The system is grounded in Work IQ, Microsoft's intelligence layer that understands organizational information, instruments, and construction. The shift strikes Copilot from a sidebar helper to an orchestrator of autonomous brokers.
Apple can be making ready a revamped Siri for WWDC 2026 that can act as an "always-on agent" able to dealing with duties throughout apps utilizing private information. Google's Gemini fashions will assist energy the upgraded Siri by way of a multi-year deal reportedly costing Apple round $1 billion per 12 months.
The convergence is unmistakable: each main platform is transferring from assistants that discuss to brokers that act. However every is approaching the issue otherwise. OpenAI's agent operates primarily by way of a browser. Anthropic's works straight on a person's desktop. Microsoft's is tightly certain to the Workplace 365 ecosystem. Apple's emphasizes on-device processing and privateness. Google's strategy with Spark is distinctive in its wager on cloud persistence and deep integration with its personal providers.
Somewhat than controlling a person's display pixel by pixel, Spark works by way of structured integrations — Google's personal Workspace APIs, and more and more, third-party connections by way of MCP. The benefit is reliability and pace: structured device use is way extra predictable than screen-reading. The drawback is that Spark, no less than initially, can solely act inside the techniques it's been linked to.
The AI mannequin behind Spark processes trillions of tokens a day — and Google says it might save enterprises billions
Spark's capabilities are inseparable from the mannequin that drives it. Gemini 3.5 Flash, additionally introduced Monday, is Google's new workhorse AI mannequin — designed particularly for the calls for of agentic workflows.
The efficiency claims are necessary. Google says 3.5 Flash outperforms its earlier frontier mannequin, Gemini 3.1 Pro, throughout practically all benchmarks, whereas operating 4 instances sooner than comparable frontier fashions when it comes to output tokens per second. An much more optimized model, obtainable inside Google's Antigravity improvement platform, runs twelve instances sooner.
Pichai framed the economics bluntly. Corporations processing roughly one trillion tokens per day on Google Cloud — a determine he stated high enterprise prospects are hitting — might save over $1 billion yearly by shifting 80% of their workloads to a mixture of Flash and frontier fashions like 3.5 Professional. In a market the place, as Pichai famous, CIOs are already "blowing by way of their annual token budgets and it's solely Could," the associated fee argument could matter as a lot as the aptitude argument.
Internally, Google's personal builders have been consuming Gemini 3.5 Flash at a staggering and quickly accelerating tempo. In March, Google was processing about half a trillion tokens per day internally. That determine has since grown to greater than three trillion — doubling roughly each few weeks. Pichai described this as a "highly effective suggestions loop" that regularly improves the mannequin.
Koray Kavukcuoglu, CTO of Google DeepMind and Chief AI Architect for Google, stated the mannequin's pace is what makes agentic use circumstances sensible. "3.5 Flash is particularly good when deploying a number of brokers concurrently and finishing long-running duties," he stated in the course of the briefing, including that Google had efficiently examined brokers constructing "a working working system totally from scratch."
The three.5 Professional mannequin, the extra highly effective sibling, is at the moment being examined internally and can roll out subsequent month.
What Gemini Spark prices and the place it suits in Google's new subscription tiers
Gemini Spark will likely be obtainable to Google AI Ultra subscribers. The corporate is concurrently restructuring its subscription tiers to make the know-how extra accessible. A brand new Extremely plan at $100 per thirty days gives a 5x increased utilization restrict than the Professional plan, together with precedence entry to Antigravity and 20TB of cloud storage. The highest-tier Extremely plan drops from $250 to $200 per thirty days, with a 20x increased utilization restrict and entry to the total suite of capabilities.
Each tiers embrace Gemini Spark, the Every day Temporary agent — a proactive morning digest that triages e-mail, calendar, and duties in a single day — and entry to the brand new Gemini Omni and three.5 Flash fashions. The pricing positions Spark as a premium product — costlier than Anthropic's Claude Professional at $20 per thirty days, however similar to the upper tiers of competing merchandise like Claude Max ($100–$200/month) and OpenAI's ChatGPT Pro ($200/month).
Why privateness, reliability, and ecosystem lock-in might undermine Google's agent ambitions
The dangers are actual and multidimensional.
Reliability stays the business's best problem. Even one of the best AI fashions hallucinate, misread directions, and make errors {that a} human would by no means make. An agent that drafts an e-mail to the flawed individual, misreads a spreadsheet determine, or sends a fee to the flawed service provider might create penalties which might be troublesome to reverse. Google's strategy of requiring express approval for high-stakes actions like spending cash or sending emails is a wise safeguard — however it additionally limits how autonomous the agent can really be. An agent that asks for affirmation at each flip isn't a lot of an agent in any respect.
Privateness is one other concern. Spark's means to synthesize info throughout a person's complete Gmail inbox, calendar, paperwork, and chat historical past means it has an awfully deep view of an individual's digital life. Google says Spark operates on a completely managed, safe runtime with remoted ephemeral digital machines, encrypted credentials, and Information Loss Prevention insurance policies. However the focus of non-public context in a single AI system — accessible by way of pure language — creates a floor space that can entice scrutiny from regulators, privateness advocates, and safety researchers.
Market timing is unsure, too. The buyer urge for food for always-on AI brokers is unproven at scale. Google says the Gemini app has 900 million month-to-month customers, however it's unclear what number of of these customers are prepared for the conceptual leap from "ask a query, get a solution" to "delegate a process, belief the end result." The historical past of digital assistants — from Clippy to early Siri to Alexa — is affected by merchandise that promised proactive intelligence and delivered frustration.
After which there may be the query of ecosystem lock-in. Spark works finest inside Google's personal providers. Whereas MCP connections to third-party apps will broaden its attain, the preliminary expertise is considered one of deep Workspace integration. For the billions of people that stay inside Google's ecosystem, this can be a pure match. For individuals who cut up their digital lives throughout Microsoft, Apple, and different platforms, Spark's utility will likely be extra restricted — no less than initially.
Woodward acknowledged as a lot when requested whether or not Spark would stay confined to the Google ecosystem. "It's going to be cross-platform in two methods," he stated — by way of MCP integrations with third-party apps, and thru availability on the internet, Android, and iOS, with duties syncing throughout gadgets through the cloud.
The actual take a look at for Gemini Spark isn't whether or not it may possibly do the work — it's whether or not individuals will let it
Google's wager with Gemini Spark is that the AI business's heart of gravity is shifting from fashions that suppose to techniques that act — and that the corporate finest positioned to win that transition is the one with probably the most complete set of client providers to behave inside. It’s a wager backed by monumental infrastructure funding. Google expects to spend roughly $180 to $190 billion in capital expenditure this 12 months — roughly six instances what it spent in 2022 — a lot of it on the AI compute required to run brokers like Spark at scale for a whole lot of thousands and thousands of customers.
The know-how, in different phrases, is arriving. The fashions are quick sufficient, the integrations deep sufficient, the fee rails safe sufficient. Google has constructed a system that may draft your emails, manage your calendar, monitor your inbox, and shortly sufficient, spend your cash — all when you sleep.
However the hardest drawback in synthetic intelligence has by no means been making a machine succesful. It has been making a human comfy. For twenty years, Google's core promise has been ten blue hyperlinks and a search field — a transaction constructed on the idea that the person is in management. Gemini Spark asks customers to renegotiate that relationship totally, at hand a set of keys to a system that’s sensible, tireless, and nonetheless, by its maker's personal admission, finest in comparison with a young person with a debit card.
Gemini Spark rolls out to trusted testers this week, with a broader beta for U.S. Google AI Extremely subscribers anticipated subsequent week.