Google used its I/O 2026 keynote on Tuesday to introduce Gemini Spark, an agentic assistant designed to run continuously in the background and complete tasks on a user's behalf. The product sits on Gemini 3.5 and Google's Antigravity agent harness, and arrives in beta for Google AI Ultra subscribers in the United States next week.

What separates Spark from the existing Gemini app is its address. Each instance is assigned a dedicated Gmail account, so users can email instructions the way they would email a person. Once instructed, Spark browses the web through Chrome, draws on the user's documents, slides and inbox, and produces drafts without an active session. The agent runs on dedicated virtual machines hosted by Google Cloud, so long jobs continue when the laptop is closed.

A pricing reset

Spark's launch coincides with a sharp cut to the AI Ultra subscription tier. Google lowered the price from $250 to $100 a month, a change announced at the same keynote according to CNBC. Subscribers get five times the usage allowance of AI Pro, twenty terabytes of cloud storage and YouTube Premium thrown in.

The cut suggests Google has concluded that broader distribution matters more than premium pricing. OpenAI, Anthropic and Microsoft are all pushing agentic products into both enterprise and consumer markets, and a $250 ceiling looked likely to leave Spark in a narrow band of power users.

Reach and openness

Google also reported that the Gemini app has crossed 900 million monthly active users, more than double the 400 million figure quoted a year earlier. That puts Spark in front of an unusually large early audience for an agentic product.

Spark supports the Model Context Protocol, the open standard for connecting language models to external tools. The choice matters. Spark can in principle reach beyond Google's own apps and operate on services its competitors expose as MCP servers. A flagship consumer agent shipping with MCP support is a small signal that the industry is settling on a shared plumbing layer.

What is still uncertain

The real question is whether Spark earns enough trust to be handed consequential work. Background agents that browse the web, read inboxes and write drafts at scale carry exactly the kind of risks federal cybersecurity agencies have been flagging in recent guidance. Google says human approval gates are available for sensitive actions, and the proof of that claim will come in months of daily use rather than in a keynote.

Sources

  1. i. techcrunch.com
  2. ii. www.cnbc.com
  3. iii. www.engadget.com
  4. iv. thenextweb.com
  5. v. tech.yahoo.com

Commentarii · 0

Add · a · Comment