Google Gemini Spark: The $100/Month 24/7 AI Agent, Explained

The era of session-bound chatbots that wait for your next prompt is officially over. At its I/O 2026 developer conference, Google unveiled Gemini Spark, a first-of-its-kind personal AI agent that lives in the cloud and works on your behalf 24 hours a day. Unlike Google's earlier experimental browser agents, Spark is a persistent, always-on assistant that doesn't depend on your laptop being open or your phone being awake to get things done .

Announced by Sundar Pichai during the opening keynote on May 19, 2026, Spark represents Google's most aggressive move into agentic AI for consumers, bundling a cloud-resident worker into a restructured $99.99-per-month subscription .

What Gemini Spark Actually Is

At its core, Gemini Spark is an autonomous software agent that runs on a dedicated Google Cloud virtual machine. This is the fundamental shift that separates it from a standard AI chatbot. Because Spark's brain lives on Google's servers, it continues to research, book, draft, and monitor tasks even when every device you own is turned off. When it finishes, it sends you a notification with the results .

Google describes the agent as acting "under your direction," meaning it operates with pre-granted permissions to access your data and take actions, but it does so as a persistent background process rather than a tool you have to manually prompt for every step .

The Technical Stack

Powering Spark is Gemini 3.5 Flash, Google's latest model optimized for fast inference and agentic workloads. Google says Flash delivers cutting-edge agent performance at less than half the price of comparable frontier models .

The agent is built on Antigravity 2.0, an internal development framework designed for orchestrating multi-step agent behaviors, and it connects to the outside world through the Model Context Protocol (MCP), a standard for linking AI models to external tools and data sources .

Launch Timeline and Pricing

The road to a public Spark was surprisingly swift after the I/O announcement.

• May 19, 2026: Unveiled at Google I/O 2026; immediate rollout to a small set of trusted testers .
• Week of May 25, 2026: Beta access opened to all Google AI Ultra subscribers in the U.S. .

The product is currently U.S. only, with Google confirming that international expansion is planned but providing no specific timeline .

Spark is not a standalone product; it is bundled exclusively with the Google AI Ultra subscription plan. In a significant pricing move designed to accelerate adoption, Google slashed the entry-level Ultra tier from $249.99 to $99.99 per month at I/O. A higher $199.99/month tier also exists with additional benefits .

The $99.99 plan also includes 5x higher Gemini app usage limits compared to the $20 AI Pro tier, 20 TB of cloud storage, and a YouTube Premium subscription .

What Spark Can Do

Spark's capabilities are a major leap beyond Google's earlier, browser-only Project Mariner. The agent is designed to handle complex, long-running workflows across multiple services .

Native Google Ecosystem Integration

Out of the gate, Spark can read, summarize, draft, and monitor emails in Gmail; manage and create Calendar events; search and organize Drive files; and work within Docs, Sheets, and Slides. It also has deep integration with Chrome, Maps, and YouTube .

Third-Party App Access

A critical upgrade from past Google agents is Spark's ability to act inside third-party services, not just Google's own products. Using MCP, Spark can interact with airline booking sites, e-commerce platforms, hotel reservation services, and other productivity apps on a user's behalf .

Autonomous Multitasking

Spark can juggle multiple tasks simultaneously. While the exact concurrency ceiling hasn't been publicly specified, reports and demos show it managing several parallel workflows—for example, researching flight options while drafting an email and monitoring a calendar for conflicts .

Permissions and Safety

Every autonomous agent raises questions about what it's allowed to do without a human in the loop. Google's approach with Spark is permission-based but broad. Users grant the agent persistent access to specific domains, and Google states that the assistant works "under your direction" .

Early pre-release documentation contained a controversial clause about the potential for autonomous purchases without explicit confirmation. By the time of the public beta launch, Google had softened this language but did not entirely remove the concept. The current implementation requires user authorization for sensitive actions like purchases, data sharing, and account changes .

Reports indicate that as of the May 2026 beta, the payments feature is not yet live, and the human approval gateway serves as the primary safety control .

From a Leaked Extension to a 24/7 Agent

Gemini Spark didn't emerge from a vacuum. It is the culmination of an 18-month journey that began with a product slip-up, evolved through a publicly named research prototype, and finally emerged as Google's flagship consumer agent .

• November 2024: A prototype Chrome extension titled "Project Jarvis" is accidentally published to the Chrome Web Store. Described as a "helpful companion that surfs the web with you," it is quickly pulled but confirms Google's agent ambitions .
• December 11, 2024: Google officially announces Gemini 2.0 and formally previews the renamed Project Mariner, an experimental Chrome extension that can navigate websites and perform actions on a user's behalf .
• May 20, 2025: Project Mariner rolls out to U.S. Google AI Ultra subscribers at the then-price of $249.99/month, with capabilities integrated into the Gemini app and Google Search .
• May 4, 2026: Google quietly shuts down Project Mariner as a standalone product. The official notice states that its technology "has been taken over by other Google products" .
• May 19, 2026: Gemini Spark is announced as the production successor, moving far beyond Mariner's browser-only limitations into a full 24/7 cloud-resident agent .

Internally, the project also went by the codename "Remy" before receiving the Gemini Spark brand .

The Agent Arms Race

Gemini Spark didn't land in an empty market. By mid-2026, the three leading AI labs have each placed a distinct architectural bet on what a personal AI agent should look like .

• OpenAI's Operator: Priced at $200/month, Operator takes a desktop-native approach, especially after the April 2026 launch of Codex Background Computer Use, which pushes OpenAI into macOS-first automation with parallel agent sessions .
• Anthropic's Claude Cowork + Skills: Reached general availability on April 9, 2026. It is browser-based with a "Managed Agents" enterprise governance layer, emphasizing safety and a portable computer-use model that can directly manipulate screens, clicks, and keystrokes .

Google's play with Spark is to anchor the agent in the cloud and tie it inextricably to its vast ecosystem of services, betting that deep Chrome and Workspace integration, zero-friction Google authentication, and a more affordable price point will make a 24/7 cloud agent more useful than a desktop app that still depends on local computing resources .

The autonomous AI assistant market is no longer theoretical. It's here, it's subscription-gated, and it's pushing users to ask a new question: not "what can I ask an AI to do?" but "what am I comfortable letting an AI do for me while I'm away?"