Google used I/O 2026 on May 19 to launch Gemini Spark, a 24/7 cloud-resident agent positioned head-to-head against Anthropic's Claude Cowork and OpenAI's ChatGPT Agent. The launch was paired with a pricing reset: Google AI Ultra now starts at $100 per month, with the existing top tier dropping from $250 to $200. Both tiers include Spark access at U.S. launch.
The agent stack: Gemini 3.5 Flash on Antigravity 2.0
Spark is built on the freshly minted Gemini 3.5 Flash model and Google's Antigravity 2.0 agentic harness, both of which Google made globally available at the conference. Each Spark instance runs on a dedicated virtual machine on Google Cloud, which is what makes the "keep working after you close the laptop" pitch technically credible — tasks survive client disconnects because the agent's session state lives on Google infrastructure, not the device.
The harness supports MCP (Model Context Protocol) connections, putting Google on the same interop standard Anthropic introduced and OpenAI has since adopted. Initial third-party integrations include Canva, OpenTable, and Instacart, alongside first-party hooks into Gmail, Docs, Sheets, Slides, Chrome, and Android's Halo system for tracking long-running jobs from a phone.
Pricing pressure on the $200/month bracket
The restructured AI Ultra lineup is the more immediate signal for buyers. Until this week, the top consumer tier sat at $250. With OpenAI's ChatGPT Pro and Anthropic's Claude Max both pinned at $200, Google's move to a $100 entry point with Spark included undercuts the entire bracket — and the $200 tier now buys a 20x usage uplift in both the Gemini app and Antigravity rather than serving as the only door to advanced features.
Google is also moving the Gemini app away from daily prompt limits toward a compute-used accounting model, where a coding or video-edit task consumes more of the monthly allowance than a short text exchange. That structurally aligns consumer pricing with API economics and lays the groundwork for selling Spark-style agents at the seat level into Workspace.
What changes for builders
Gmail is the most consequential surface in the launch. Spark gets its own email address, can monitor an inbox for customer questions, and can draft replies grounded in Docs and Sheets — effectively shipping a default agentic CRM layer to anyone with a Workspace account. VP Josh Woodward called out small businesses already using Spark to triage inbound mail as the early validation case.
For teams building on Gemini, the practical implication is that Antigravity 2.0 is now the assumed substrate for production agents on Google's stack, with MCP as the external tool interface and dedicated-VM execution as the runtime contract. Spark functionality is slated to roll out to the existing Gemini macOS app later this summer.
