Gemini Spark review — Google's always-on, 24/7 personal AI agent
Contents
Gemini Spark review — Google's always-on, 24/7 personal AI agent (Google I/O 2026)
Gemini Spark is the '24/7 personal AI agent' unveiled at Google I/O 2026. It monitors Gmail, Calendar, and Workspace in the background and keeps working on a cloud VM even when your device is off. This article covers Spark's features, pricing (the keynote price and the current official page price differ), how it works, a comparison with competitors (ChatGPT Pulse, Claude Max), and the key points Korean users need to know right now.
⚠️ Important notice (as of May 20, 2026): The Google I/O 2026 keynote announced AI Ultra in two tiers, $99.99 and $199.99, but at the time of writing the official subscription page (gemini.google/subscriptions/) shows a single AI Ultra tier at $249.99. The new pricing appears to be rolling out gradually, starting with some US users, and may change until it is reflected on the official page, so we recommend rechecking the official page right before subscribing.
Basic information
| Item | Details |
|---|---|
| Developer | Google (Google DeepMind) |
| Name | Gemini Spark (keynote name) — consolidated and expanded as "Gemini Agent" in official listings |
| Announced | May 19, 2026 (Google I/O 2026 keynote) |
| Release stage | Trusted testers first → beta for US AI Ultra subscribers |
| Price | Keynote announcement: Included in AI Ultra at $99.99 or $199.99 Official page (as of 2026.5.20): Single AI Ultra tier at $249.99 Korea: AI Ultra ₩360,000/month |
| Official site | gemini.google.com / gemini.google/overview/agent/ |
| Category | Personal AI agent (agentic AI / personal assistant) |
| Underlying model and infrastructure | Gemini 3.5 Flash + Antigravity 2.0 + dedicated Google Cloud VM |
| Supported platforms | Android, iPhone, Gemini Mac app (this summer), Chrome (this summer) |
| Available regions | US only, English only, ages 18 and up (as of May 2026) |
Key features
- Always-on (24/7) background operation: Because it runs on a dedicated VM in Google Cloud, work continues even when your device is off or the laptop lid is closed. It notifies you when the task is done.
- Multi-input task instructions by voice and text: The keynote demo showed it composing and sending email, coordinating schedules, and handling multi-step scheduling by voice.
- Workspace and external service integration: It connects information and automates tasks across Gmail, Docs, Sheets, Calendar, Tasks, Drive, Keep, Google Maps, YouTube, and more. Users can choose and manage which apps to connect.
- Use cases: Detecting hidden subscription charges in credit card statements, tracking email updates from your child's school, gathering project notes from Gmail into Docs, multi-step travel and event planning, online reservations and purchases, and more.
- Confirmation for sensitive actions: Irreversible actions such as sending email, making payments, and adding calendar entries ask the user for confirmation before executing.
- 'Labs' experimental stage: Per the official guidance, Gemini Agent is classified as an 'experimental feature in early development', and Google states that 'user supervision is important to prevent unintended behavior'.
Plan comparison
Gemini Spark/Agent is not sold on its own; only Google AI Ultra subscribers in the UScan use it. Since the price differs between the keynote announcement (2026.5.19) and the official page (checked 2026.5.20), both are listed.
New pricing structure announced at the keynote (rolling out gradually)
| Plan | Price | Spark/Agent | Key benefits |
|---|---|---|---|
| Free | $0 | ❌ | Basic Gemini app, access to 3 Flash |
| AI Plus | $7.99/month | ❌ | Higher limits than Free, 200 GB storage, limited access to Veo 3.1 Lite |
| AI Pro | $19.99/month | ❌ | 5 TB storage, 1M-token context, 5x NotebookLM limits, Antigravity access |
| AI Ultra $100 | $99.99/month | ✅ (US beta) | 5x the usage limits of AI Pro, 20 TB storage, YouTube Premium, priority Antigravity access |
| AI Ultra $200 | $199.99/month | ✅ (US beta) | 20x the usage limits of AI Pro, global access to Project Genie, Deep Think |
Current prices shown on the official page (checked 2026.5.20)
| Plan | US | Korea |
|---|---|---|
| Free | $0 | ₩0 |
| AI Plus | $7.99 | ₩11,000 |
| AI Pro | $19.99 | — |
| AI Ultra | $249.99 (single tier) | ₩360,000 (single tier) |
? Note: In the same announcement, Google said it is abolishing daily prompt limits and switching to 'compute-used' billing (based on computing resources consumed). When the limit is exceeded, it automatically downgrades to Gemini 3.5 Flash, and AI Pro and Ultra subscribers can buy separate pay-as-you-go credits. This change, too, is in gradual rollout.
Pros
- The 'assistant → agent' paradigm made real: A structure where work continues on a cloud VM without the user keeping an app open sets it apart from the 'conversation session' model of ChatGPT and Claude.
- Deep integration with the Google ecosystem: It delivers immediate value to users whose everyday infrastructure, Gmail, Calendar, Docs, Sheets, Maps, YouTube, already lives on Google.
- Confirmation step for sensitive actions: Applying human-in-the-loop by default to irreversible actions partly mitigates the risk of a runaway agent.
- Multiple starting points: Tasks can be given by voice, text, or image input, and continued across mobile and desktop.
Cons / drawbacks
- Starts US-only, English-only, adults-only: No clear launch schedule has been announced for most global markets, including Korea. Workspace accounts, student accounts, and those under 18 cannot use it.
- Labeled an 'experimental feature': Google's official guidance stresses that it is 'an experimental feature in early development requiring user supervision', so it is some way from a proven, stable product.
- Pricing confusion: The keynote price ($99.99/$199.99) and the official page price ($249.99) differ, and the new pricing has not been applied in some markets, including Korea. Confirming the exact price before subscribing is essential.
- Privacy concerns in the US market where it is in use:
- An onboarding screen leaked just before the keynote included the notice 'Spark may share information or make purchases without asking you', which caused controversy. No Spark-specific privacy policy had been published separately as of the keynote.
- The class action Thele v. Google LLC (Case No. 5:25-cv-09704) was filed in November 2025 and was in the discovery stage as of March 2026. Google has filed a motion to dismiss. Since Spark accesses Workspace data, the future course of this lawsuit is a variable that will affect trust.
- In a separate case, Rodriguez v. Google LLC, a jury returned a $425.7 million damages verdict in September 2025 (covering about 98 million Gmail users), and Google is appealing.
- The double edge of 'compute-used' billing: Users who do a lot of video editing and agent work may burn through their limits faster than under the old daily limit system and need to buy additional pay-as-you-go credits.
- Not available on its own: Even if you only want Spark, you need at least an AI Ultra subscription ($99.99–$249.99).
Recommended for
- ✅ US-based heavy Google Workspace users: Office workers and freelancers who spend a lot of time in Gmail, Docs, and Calendar
- ✅ Solo business owners and creators who need to automate repetitive work
- ✅ Developers and tech leads: The same AI Ultra plan also gives priority access to Antigravity
- ❌ Residents outside the US, including Korea: Global rollout schedule undecided (hard to use directly for now)
- ❌ Users sensitive to privacy and data control: Recommended to decide after the Spark-specific privacy policy is published and the Thele lawsuit's progress is clear
- ❌ Occasional AI users: A monthly spend of at least $99.99–$249.99 is a burden
Competitor comparison
| Item | Gemini Spark/Agent | ChatGPT Pulse | Claude (Max) |
|---|---|---|---|
| Developer | OpenAI | Anthropic | |
| Underlying model | Gemini 3.5 Flash + Antigravity 2.0 | GPT-5.5 (flagship as of April 2026) | Claude Opus 4.7 / Sonnet 4.6 |
| Core concept | Task-executing agent (runs email, scheduling, reservations) | Information curation (personalized morning briefing cards) | Powerful reasoning, coding, and document analysis, with Agent Teams support |
| Execution model | Cloud VM, works even with the device off | Background research overnight, delivered as cards in the morning | Agent work through Claude Code and others (user session based) |
| Entry price (monthly) | AI Ultra $99.99–$249.99 (pricing rolling out gradually) | ChatGPT Pro $200 (for Pulse) — the separate $100 Pro plan centers on other features such as Codex | Claude Max $100 (5x) / $200 (20x) |
| Workspace and email integration | Native with Gmail, Docs, Calendar, etc. | Connected via connectors for Calendar, Gmail, etc. (opt-in) | Can integrate via MCP and connectors |
| Korean support | ❌ (English only) | ✅ | ✅ |
| Available regions | US only | US first, expanding gradually | Global |
? Positioning summary: Even among 'background AI' products, the characters differ clearly. ChatGPT Pulse is an information curation tool for 'receiving a morning briefing card', while Gemini Spark is an action agent that 'actually does the work for you'. Claude offers powerful reasoning, coding, and agent capabilities, but its consumer workspace integration is relatively weak. The three tools are less direct competitors than complements aimed at different usage patterns.
Verdict
Gemini Spark is the product with which Google declares its paradigm shift 'from assistant to agent'. The always-on cloud VM structure and deep integration with everyday infrastructure such as Gmail, Calendar, and Docs are an approach only Google can attempt.
At launch, however, it comes with the clear limitation of being US-only and English-only, plus the confusion of a mismatch between the keynote price and the official page price. The absence of a privacy policy and the ongoing Thele class action are trust variables as well. For Korean users, rather than rushing to subscribe now, it is more sensible to monitor developments while watching for the global rollout schedule, Korean-language support, and stabilization of pricing. At the same time, if you sort out in advance which tasks in your workflow 'would be nice for an AI agent to handle in the background 24 hours a day', you will likely extract value faster when you can actually use it.
Rating: ⭐⭐⭐⭐☆ (4.0/5)
A rating of potential. Re-evaluation needed once it is available in Korea, the privacy policy is published, and pricing stabilizes.
Contents
Related posts
Claude AI review — ChatGPT's real rival, an AI chatbot strong at writing and document work
A hands-on look at Claude, the AI chatbot from Anthropic: core features, pricing, pros and cons, and Korean-language performance. An honest take on how it differs from ChatGPT and Gemini and who it suits best.
The complete guide to Claude Code Agents — from subagents to Agent View (2026)
Anthropic's Claude Code is evolving beyond a simple coding assistant into a multi-agent development platform. Subagents, Agent View, and Agent Teams — a systematic rundown of Claude Code's agent features.
Cursor, Genspark, Firefly vs. ChatGPT, Claude, Gemini — multi-model wrappers or first-party AI: which should you use? (2026)
The AI market is splitting into two camps: the 'first-party' camp, where OpenAI, Anthropic, and Google compete directly with their own models, and the 'multi-model wrapper' camp, which combines those models into experiences specialized for coding, research, and design. As of 2026, the growth of wrapper services has begun to threaten the first parties. So which should you choose, and when?
Claude Design review — Anthropic's new AI design tool: a threat to Figma and Canva?
A look at Claude Design, the new product Anthropic unveiled on April 17, 2026. This tool creates prototypes, slides, and landing pages from text prompts and conversation alone. We cover what it means for designers and non-designers alike, plus pricing, key features, pros and cons, and a comparison with competing tools.