Google Launches Personal AI Agent Gemini Spark
Google introduced the Gemini Spark agent, powered by Gemini 3.5 Flash, which continuously executes user tasks in the cloud. The agent is integrated with Gmail and Workspace, and a beta test for AI Ultra subscribers in the US will begin next week.
Gemini Spark: Why Google Is Moving Away from the Chatbot Model and What It Means for the AI Agent Market
[The Gist]: What's Really Happening
Google has just made the most radical pivot in the positioning of its AI product in its entire history. Gemini Spark is not an improved version of a chatbot, but a full-fledged replacement for the very concept of a chatbot. The point is not that the agent can answer queries while you sleep—the point is that Google is consciously abandoning the interaction model it itself popularized through its search engine. Instead of "user asks — AI answers," a new paradigm is introduced: "user delegates — AI reports completion."
The key shift is architectural: Spark runs on dedicated Google Cloud virtual machines, not on the user's device. This means the agent continues to function even when the laptop is off or the phone is locked—just send it a task via a special Gmail address. This approach transforms Gemini from an app into a personal infrastructure operating 24/7.
The technological foundation deserves special attention: Gemini 3.5 Flash, on which Spark is built, delivers up to 289 tokens per second—four times faster than its predecessor. The optimization is specifically tailored for multi-step tasks where processing speed is critical for maintaining user interest.
Timeline and Context
To understand the logic behind the launch, we need to rewind and look at the sequence of events leading up to this moment.
December 2023 – June 2024. Google launches Gemini 1.0, then 1.5 Pro—a model with a giant context window. The product is positioned as a "smart assistant" for search and generation. The model is impressive, but it remains fundamentally in the same "question-answer" paradigm as competitors. Users get used to the format, but retention leaves much to be desired—people interact with AI but don't trust it with actions.
September 2024. Anthropic releases Computer Use—a feature that allows Claude to control the cursor and perform actions in the browser. Then OpenAI launches Operator. The market sees agents capable not only of answering but also of acting for the first time. Google observes, analyzes telemetry, and understands: the next stage is not answers, but execution.
2025. Google DeepMind accelerates the development of agent architecture. Simultaneously, internal consolidation occurs: Vertex AI Agent Builder, Gemini Enterprise Agent Platform, integration with Workspace. It becomes clear that the agent layer will be overlaid on the entire ecosystem, not spun off as a separate product.
Late 2025. Rumors about an internal project codenamed "Spark" leak into the industry. Sources describe it as "Google's version of 'Lobster'"—a reference to OpenClaw, a competitor's agent product. Google remains silent, but the hiring pace in Google Labs VP Josh Woodward's team accelerates sharply.
May 2026, Google I/O. Spark is unveiled. Beta for trusted testers this week, for paid AI Ultra subscribers in the US starting next week. Simultaneously, Google radically restructures its pricing: the base AI Ultra tier drops to $100 per month instead of the previous $250, and the top tier to $200. This is not just cosmetic rebranding but a strategic move to expand the funnel of paying users ahead of the agent's mass launch.
Who Wins and Who Loses
Winners:
Google Workspace subscribers. Spark has native access to Gmail, Docs, Slides, and other ecosystem tools without additional setup. These are ready-made data sources for the agent—emails, calendars, files on Drive. Competing agents will have to connect to these services via API and OAuth, creating friction and reducing conversion. A user already living inside the Google ecosystem gets an agent that "knows" their context from day one.
Small businesses. Josh Woodward specifically highlighted this segment: Spark can monitor incoming emails, track customer requests, and ensure no message goes unanswered. For small companies without a staff of secretaries, this is equivalent to hiring a virtual employee for $100 per month—less than the cost of one working day for an assistant in the US.
Google Cloud. Each Spark instance runs on a dedicated virtual machine. This means guaranteed utilization of cloud resources tied to the number of active subscribers. Direct monetization of infrastructure through AI workloads is exactly the scenario that justifies multi-billion dollar investments in server capacity.
Losers:
Anthropic and OpenAI. Both companies have already launched personal agents, according to TechCrunch. But they lack what Google has—native integration with an email service used by over 1.8 billion people and an office suite that permeates corporate America. Spark doesn't need to convince users to connect Gmail—it's already there.
Startups in the AI automation niche. Zapier, Make, and dozens of smaller players who built businesses on connecting services to each other face a direct threat. If Google's agent can itself go through the chain of "check email → find invoice → pay it in the banking app → save receipt to Drive," the need for intermediate middleware sharply decreases.
What the Media Isn't Saying
Most publications focus on Spark's functionality—and miss one critical point: Google is consciously sacrificing its search monetization model. A chatbot with ad inserts generated revenue here and now. An agent performing tasks in the background doesn't show ads—it generates actions, not page views.
This means Google is betting on the subscription model as a replacement for advertising. Hence the aggressive price cuts: AI Ultra for $100 instead of $250. Google understands that mass adoption of agents is only possible at an affordable price. And it's ready to temporarily lose revenue per user to capture the market. The long-term bet is simple: an agent you trust with routine tasks becomes indispensable—and the subscription price can increase over time.
The second underestimated factor is the Android Halo feature. This tool allows mobile users to track Spark's actions in real time. Why is it needed? The answer is a psychological barrier. Google's research shows that the main obstacle to agent adoption is not technical limitations but the fear of losing control. The user must see what the agent is doing to start trusting it. Android Halo is not a technical but a behavioral feature. Google is designing not only code but also user habits.
Forecast: Next 30 Days and 90 Days
30 days (through end of June 2026).
The beta test among AI Ultra subscribers will reveal initial usage patterns. I expect the most popular scenarios to be: automatic email processing (filtering, summarization, responses to standard queries), monitoring financial transactions, and tracking deadlines from school and university emails—exactly the scenarios Google demonstrated in the presentation.
The key question is how stable Spark will work with Gmail. The email client will be the main battleground: if the agent starts missing emails or, worse, sending incorrect replies, trust will collapse instantly. Google understands this, so it has implemented a mandatory confirmation request before actions like sending emails or spending money.
Competitors won't sit idle. I expect an announcement from Anthropic about deep integration of Claude with Microsoft 365—a mirror response to the Spark + Workspace pairing. OpenAI will likely accelerate the launch of a similar agent with access to the Apple ecosystem.
90 days (through end of August 2026).
By this time, Spark should appear in the standalone Gemini app for macOS, where it will gain access to local user files. This is a qualitative leap: the agent will go beyond the cloud and start operating on-device data. Scenarios like "find all PDFs with invoices from the last six months, sum up expenses, and build a chart" will become reality.
The market will begin to segment. Three types of users will emerge: those who fully trust the agent with routine tasks; those who use it selectively for specific tasks; and those who are fundamentally unwilling to delegate control to AI. Google will be able to target offers to each segment—from basic monitoring to a full-fledged "digital twin."
The most important indicator is retention 30 days after activation. If more than 40% of users who activated Spark in the beta continue using it after a month, it will be a signal for mass deployment. If retention is low, Google will have to rethink the UX and possibly return to a hybrid "chatbot with agent features" model instead of a pure agent.
Numbers to watch: each paid Ultra subscriber brings in $100 to $200 per month. With even a 5% conversion rate from Gemini's 900 million user base, that's $4.5 billion in annual subscription revenue alone—excluding advertising and cloud revenue. This math explains why Google is willing to cannibalize its own search business for an agent-driven future.
— Editorial Team
No comments yet.