Google’s Gemini Spark keeps working with the lid shut — paid beta lands May 26

Gemini Spark is the first version of Google’s consumer AI assistant that does not need the screen to be on. Announced at Google I/O 2026, Spark runs on a dedicated virtual machine inside Google Cloud and keeps acting on its assigned task even after the user closes the laptop, locks the phone or signs out for the night. When the user comes back, the work has continued in the background under the same instructions, with the partial results waiting in Gmail, Docs or Drive.

For the reader who has spent two years asking a chatbot to draft an email and copying the result by hand, that is the change. Spark is the moment Google stops treating its assistant as a conversation and starts treating it as a contracted worker that holds its own session, manages its own queue and reports back when something is done. The agent is built on Gemini base models wired into an agentic harness from Google Antigravity, the runtime layer Google introduced last year to let language models call tools, browse and execute tasks without a human pressing send between each step.

Spark launches with deep hooks into the apps where most people already work. Inside Google’s own stack it can read Sheets and Docs, draft and send Gmail, file results in Drive and act across Workspace. Outside it, the first batch of partner integrations covers Canva, OpenTable and Instacart — design, restaurants and groceries — with more apps promised in the following weeks. The example Google led with at the keynote was modest and revealing: ask Spark to scan the credit card bill at the end of each month and flag hidden fees. That is not a productivity demo. That is the assistant doing the chore the user keeps not doing.

The pricing news arrived in the same breath. Google AI Ultra, the top consumer plan, was cut from $250 to $100 a month, and the old $250 tier was rebadged as a $200 plan with identical capabilities. Both tiers get Spark beta access in the United States. The $100 plan also bundles five times the Gemini usage of the $20 AI Pro tier, twenty terabytes of Drive storage and YouTube Premium. In a single move Google undercut Anthropic’s Max plan, which sits in the same monthly band, and put a personal agentic assistant inside the same monthly outlay as a streaming bundle.

Spark works across Android, iPhone and the Mac Gemini app, and Google says it can be assigned recurring tasks rather than one-shot queries — the same email summary every Monday, the same supplier-bill check every month, the same flight watch until a fare drops. Each running task is visible in a dashboard the user can pause or cancel. That dashboard is the second reason this launch matters in a way the model release announcements of the last year did not: it gives the user a single screen where the AI’s open work is listed, the way an inbox lists email. The asymmetry of giving an assistant an instruction and never seeing what it is doing is the part agentic AI keeps failing on, and Spark is the first major-vendor attempt to fix it as a product surface.

There are real caveats. Spark is US-only at launch, both for the trusted-tester wave and the Ultra subscriber beta. International availability is not on Google’s published roadmap and pricing outside the US has not been confirmed. The agent’s reach into third-party apps is still narrow — three partners at launch — and the cross-app errands Google demoed only work when the user has the right accounts connected in the right order. The first month of any agentic product is also the month its hallucinations matter most: a chatbot that imagines a fact is annoying, but an agent that imagines an OpenTable reservation books a table that does not exist. Google’s pitch is that the user is always in charge and can revoke an action mid-task; that pitch survives or breaks on what the early-access beta actually does without supervision.

On the platform layer the launch is harder to undo. Spark turns Google AI Ultra from a Gemini upgrade plan into the price of admission for an agent that lives next to Gmail. It also reframes the comparison every other AI vendor now has to answer: not which model wrote the better paragraph, but whose assistant kept working after the user went to dinner. Anthropic, OpenAI and xAI all have agentic products in some form, but none of them sit inside Gmail, Docs and Drive by default.

Trusted-tester access opened this week and Google AI Ultra subscribers in the United States get the beta the week of May 26. Samsung’s Unpacked event in July is the next major Gemini surface — the smart glasses Google previewed at I/O run on the same Spark stack — and Google has said additional Spark partner integrations will land in waves through the summer.