Gemini Spark Is Google's Bet on a 24/7 Personal AI Agent — And It Lands on Ultra Next Week

Tucked inside Google's I/O 2026 keynote today was the company's most ambitious bid yet to turn Gemini from a chatbot into a true personal agent. Gemini Spark, powered by the freshly announced Gemini 3.5 model, runs as a persistent 24/7 worker on private Google Cloud infrastructure and can complete long-running tasks while the user is offline — planning a party, keeping a spreadsheet of vendors current, drafting and re-drafting documents in response to incoming email, and firing reminders at the right moment.

The product surface is broader than a single chat thread. Spark integrates with Sheets, Gmail, Calendar, and Chrome, and Google demoed it coordinating across all four — picking up a vendor list from Sheets, pulling pricing from Gmail threads, and reconciling availability against Calendar without a human in the loop on each step. Agentic browsing through Chrome lands later in 2026, which is when the agent will be allowed to act on websites that don't expose a clean API.

Spark enters beta next week, gated initially to US Google AI Ultra subscribers. Google used the moment to recut Ultra pricing: a new $100 tier slots between the existing plans, and the top-end Ultra plan drops from $250 to $200 a month. That sets up a direct tier-for-tier confrontation with OpenAI's ChatGPT Pulse and Anthropic's Claude Cowork, both of which have spent the spring expanding always-on agentic workflows into paid subscriptions.

The technical claim Google leaned on hardest was speed. Gemini 3.5 Flash, the model behind Spark's reactive paths, delivers 289 tokens per second — about four times the throughput of competing frontier systems — and posts a 1,656 Elo on GDPval-AA, the real-world agentic benchmark where the prior Gemini 3.1 Pro scored 1,314 and trailed Claude Sonnet 4.6. Pichai said Google has been focused on three axes for the past year: "agentic coding, long-horizon tasks, and real-world workflows." Spark is the consumer-facing application of all three.

The agent push isn't isolated. At the OS layer, Android Halo, shipping summer 2026, gives Spark and third-party agents a dedicated home on the phone — a system surface rather than a chat window — while the rebranded Gemini Intelligence layer threads the model through Search, Workspace, and Chrome. Together they sketch a picture in which Spark isn't a destination so much as a layer the user invokes from wherever they already are.

The early questions are the obvious ones: trust, scope of permissions, and how the agent handles long-horizon tasks without going off-script or hallucinating side effects. Google says Spark runs in user-isolated Cloud sandboxes with action logs and per-app permission grants, but the company did not detail the safety stack on stage. With the Gemini app already at 900 million monthly active users and 9.7 trillion tokens of monthly throughput, Spark gives Google a single unified destination for agentic work — and the clearest reason yet to upgrade to Ultra rather than churn to a rival.