Google Gemini Spark: 24/7 Agentic AI Assistant Explained

TL;DR

Google announced Gemini Spark on May 19, 2026 at I/O 2026 — a 24/7 personal AI agent that runs on dedicated Google Cloud virtual machines, even when your phone is locked or your laptop is closed. It is powered by Gemini 3.5 Flash running inside the Antigravity agent harness, which Google says delivers roughly 12× the throughput of standard frontier models for agentic workloads¹². Spark integrates natively with Gmail, Docs, and Slides, plus launch-day MCP connections to Canva, OpenTable, and Instacart, with Adobe, Asana, Spotify, Uber, GitHub, Notion, and Slack to follow. It rolls out to trusted testers first and to Google AI Ultra beta subscribers in the U.S. the week of May 25, 2026. A new $100/month AI Ultra tier also launched the same day, while the top tier dropped from $250 to $200/month³. Payments are gated by Google's Agent Payments Protocol (AP2), but Google's own disclaimer warns Spark "may do things like share your info or make purchases without asking"⁴.

What You'll Learn

• What Gemini Spark actually does — and how it differs from a chatbot
• The Gemini 3.5 Flash + Antigravity stack powering Spark and its claimed 12× throughput
• Launch integrations (Gmail, Docs, Slides + MCP partners) and what's coming next
• Pricing: the new $100 Google AI Ultra tier and the $50/month price cut on the top tier
• Safety: how AP2, opt-in connectors, and "Stop / Take Control" limit risk
• How Spark stacks up against OpenAI's ChatGPT Agent, Anthropic's Claude Cowork, and the open-source OpenClaw
• Practical: how to get access, and the caveats to know first

What Is Gemini Spark?

Gemini Spark is Google's first fully agentic, 24/7 personal assistant in the Gemini app. Unlike a chatbot that waits for prompts, Spark is a persistent runtime — it stays online continuously on dedicated Google Cloud VMs and takes action across your digital workspace under your direction²⁵.

Practical examples Google demonstrated at I/O include:

• Building a list of critical deadlines buried in your Gmail and emailing it back to you each morning
• Drafting a follow-up email summarising a long thread, then leaving it in your Drafts for review
• Watching credit-card statements and flagging hidden fees every month
• Preparing for an event ("get ready for my kid's t-ball game") by adding items to your Instacart cart and booking restaurant reservations through OpenTable⁵

Three things make this different from "Gemini in Gmail" today:

1. It runs even when your device is off. The agent lives on a Google Cloud VM, not your phone.
2. It has a dedicated email address. You can email Spark like you'd email an assistant, and it triages from there⁶.
3. It can drive a browser. Spark interacts with the web directly through a remote Chrome instance you can interrupt at any time⁴.

How Spark Works: Gemini 3.5 Flash + the Antigravity Harness

Spark is built from two pieces Google introduced the same day.

Gemini 3.5 Flash is the model. It supports a 1M-token input context window and 64K output, accepts text, image, audio, and video, and has dynamic thinking on by default. On Google's own benchmark table it scores 76.2% on Terminal-Bench 2.1, 83.6% on MCP Atlas, and 84.2% on CharXiv Reasoning — outperforming Gemini 3.1 Pro on the coding and agentic suite⁷. (A Gemini 3.5 Pro model is in internal use and is expected next month.)

Antigravity is the agent harness around the model. It wraps Gemini calls with infrastructure for goal persistence, task decomposition, tool orchestration, safety constraints, and state recovery. Inside Antigravity, Google reports an optimised variant of Gemini 3.5 Flash that runs roughly 12× faster than standard frontier models with the same quality, combining speculative tool-call prediction with model compression tuned for agentic patterns¹⁸.

That throughput claim matters more than it sounds. Agentic workflows often chain dozens of model calls per task — a 12× throughput multiplier translates directly into how many parallel "background" tasks Google can give every Ultra subscriber without compute costs cratering the unit economics.

The harness is no longer Google-internal. Antigravity 2.0 ships the same day as a standalone desktop app, a CLI, an SDK, and a Managed Agents endpoint in the Gemini API — meaning developers can spin up an isolated Linux environment with one API call and run their own agents on the same runtime that powers Spark⁹¹⁰.

Launch Integrations

Sources for the integration list: Google's I/O blog post², TechCrunch's I/O coverage⁶, and Engadget's launch piece⁵.

The MCP-based design matters: because Spark talks to partners through the Model Context Protocol, the integration list scales without Google writing custom connectors for every service.

Pricing: The New Google AI Ultra Tiers

Spark requires a Google AI Ultra subscription. Google restructured AI Ultra the same day to make room for it³:

⚠ Prices change frequently. The values above are for illustration only and may be out of date. Always verify current pricing directly with the provider before making cost decisions: Anthropic · OpenAI · Google Gemini · Google Vertex AI · AWS Bedrock · Azure OpenAI · Mistral · Cohere · Together AI · DeepSeek · Groq · Fireworks AI · Perplexity · xAI · Cursor · GitHub Copilot · Windsurf.

The $100 tier also includes 20 TB of cloud storage and a YouTube Premium individual plan. The simultaneous $50/month price cut on the existing top tier ($250 → $200) is a clear signal Google wants to keep heavy agentic users from defecting to OpenAI's $200 ChatGPT Pro plan³.

Spark itself is available to AI Ultra subscribers in the U.S. starting the week of May 25, 2026, after a brief trusted-tester period².

Safety and User Control

Persistent agents that can spend money and send mail on your behalf are a fundamentally different risk surface than a chatbot. Google built several controls — and a frank disclaimer.

Permission gating. For high-risk actions like spending money or sending email autonomously, Spark is "generally supposed to" ask first. The word "generally" is doing a lot of work in Google's own help text⁴.

Agent Payments Protocol (AP2). Payments flow through AP2, an open protocol Google first announced in September 2025 with 60+ partners — including PayPal, Mastercard, American Express, Adyen, Coinbase, Salesforce, and Worldpay. AP2 uses cryptographically signed "Mandates" — verifiable digital credentials that record specific user instructions (Intent Mandates set the goal, Cart Mandates confirm the final purchase). The latest revision is v0.2.0 from April 2026, with reference implementations in Python, TypeScript, Kotlin, and Go¹¹.

Opt-in connectors. Every connected app is opt-in. Users can revoke any app, delete Gemini activity, and stop or take control of the remote browser at any time⁴.

The disclaimer. Google's own help text states Spark "may do things like share your info or make purchases without asking" and recommends supervision⁴. That is unusually direct language from a Google product launch, and it sets expectations: this is a beta-quality agent for early adopters, not a fire-and-forget productivity tool for novices.

Spark vs. OpenAI vs. Anthropic vs. OpenClaw

Here's where Spark sits in the May 2026 agent landscape:

A few honest takeaways:

• Spark's edge for consumers is the off-device runtime. Consumer ChatGPT Agent needs an active session; OpenClaw needs your laptop on. OpenAI's Workspace Agents do run persistently in the cloud — but only on business plans (Business, Enterprise, Edu, Teachers)¹³. Spark brings 24/7 persistence to individual consumer subscribers.
• OpenClaw's edge is control. Self-hosted, transparent, no subscription — it just needs API credits.
• ChatGPT Agent's edge is screenshot-grounded browsing — useful for sites without MCP support, slower than MCP-native calls.
• Claude Cowork's edge is depth on knowledge work — Anthropic's models tend to lead on long-context synthesis.

None of these is "better" — they're optimised for different jobs.

Daily Brief: Spark's Lighter Sibling

If you don't want a full background agent, Google also shipped Daily Brief the same day. It's a personalised morning digest pulling from Gmail, Calendar, and Tasks — closer to a "smart summary" than an autonomous agent, with no off-device runtime. Daily Brief is available to AI Plus, Pro, and Ultra subscribers in the U.S., not gated behind Ultra alone¹⁶. For most users, it's the cheaper, lower-risk on-ramp to agentic features.

What This Means for AI Agents in 2026

A few patterns to watch over the rest of 2026:

1. Off-device persistence is becoming the default. Spark, OpenAI's Workspace Agents, and Anthropic's June 15, 2026 "programmatic credit pool" all push the work off the user's laptop and onto vendor-managed runtimes. The era of "your agent runs while your laptop's open" is ending.
2. MCP is the connector substrate. Spark, Claude Cowork, and OpenClaw all rely on MCP for third-party integrations. Vendors that ship MCP servers are betting on agent-driven traffic.
3. Agentic payments need their own protocol. AP2 is Google's bet, but expect Anthropic, Stripe, and the EMV/ISO bodies to push competing or complementary specs.
4. Pricing is consolidating around $100–$200/month. Google AI Ultra now starts at $100, matching OpenAI's lower ChatGPT Pro tier; Google's $200 top tier matches ChatGPT Pro's premium price. Casual chatbot users stay free; agent users pay.
5. Disclaimers will get more honest. Google's blunt "may do things … without asking" language is closer to a software-EULA reality check than a marketing promise. Expect competitors to follow.

For developers, the more important launch may actually be Antigravity 2.0 — the same harness, available as desktop app, CLI, SDK, and Managed Agents API. Spark is the marquee app; Antigravity is the platform other apps will be built on. Our take on the broader agent shift is in AI Agents: The Next Frontier in Software Development, and the underlying computer-use trajectory is covered in GPT-5.4 Beats Humans at Computer Use.

How to Get Access

If you're a developer rather than an end user, the better entry point is Antigravity 2.0 (desktop app, CLI, SDK) or the Managed Agents endpoint in the Gemini API — both available now in AI Studio⁹¹⁰.

Bottom Line

Gemini Spark is the most aggressive bet yet that the productivity agent runs in the cloud, not on your laptop. The combination of Gemini 3.5 Flash, the Antigravity harness's 12× throughput claim, an off-device VM runtime, and a dedicated Gmail interface is novel — and the pricing structure (a new $100 AI Ultra tier plus a $50/month price cut at the top end) shows Google is willing to discount to win agent share.

But the rollout is staged carefully: trusted testers first, U.S.-only Ultra beta next week, partner MCP rollouts spread across summer. And Google's own help text concedes Spark "may do things … without asking." For early adopters that's an acceptable trade; for everyone else, Daily Brief is the safer place to start.

The bigger story for developers is Antigravity 2.0 — the same agent runtime, now opened up as desktop app, CLI, SDK, and Managed Agents API. Spark is the consumer surface; Antigravity is the platform every other agent on Google Cloud will be built on.