Google used the opening keynote of I/O 2026 on May 19 to ship Gemini Spark, a 24/7 cloud agent that lives inside Gmail, Docs, Calendar, and a growing list of partner apps. Spark is built on Gemini 3.5, runs without a tab or device open, and reaches out for confirmation before any action that spends money or touches another person’s inbox. Google also cut the monthly price of AI Ultra from $250 to $200 and rolled Spark into the existing $100/month AI Pro+ tier. Spark beta opens to trusted testers this week and to US AI Ultra subscribers next week, with broader access through the summer.
What’s actually new about Gemini Spark
Gemini Spark is Google’s first always-on, cross-app agent in the consumer Gemini product. Prior Gemini features lived inside individual apps (Gmail Help me write, Docs summarize, Sheets formula assist). Spark is a single agent that reasons across Gmail, Docs, Slides, Sheets, Calendar, Drive, and any connected third-party app. It runs in Google’s cloud so your laptop or phone does not need to be on, and it has its own designated Gmail address — you can email Spark a task the same way you’d email a colleague.
The architecture is meaningful. Spark is not a chat sidebar that calls an API; it is a persistent agent with memory, context across sessions, and the ability to spawn sub-agents for parallel work. Google demonstrated Spark monitoring a credit card statement for hidden subscriptions, pulling project notes from a quarter’s worth of Gmail threads into a fresh Doc, tracking emailed updates from a child’s school across the calendar, and confirming purchases via OpenTable and Instacart before executing. Confirmation gates are mandatory for spend, sends, and calendar pads — the model never silently transacts.
Why it matters for AI agents in 2026
- Consumer agent space is real. Until now, mainstream consumer AI has been chat. Spark is the first major consumer agent with cross-app context, persistent memory, and partner integrations shipped to a wide tier.
- Workspace becomes Google’s agent platform. Spark is the first surface where Google’s productivity suite is exposed to an agent natively — no MCP server, no plugin, no separate auth flow.
- $100 entry price reshapes the agent tier. Anthropic Claude for Small Business sits at $30-50/user. ChatGPT Pro is $200. Spark in AI Pro+ at $100 is positioned between, with bundled storage, YouTube Premium, and 5x Gemini usage.
- AI Ultra cut to $200 signals price compression. The $50 reduction is the first downward move at the top tier since ChatGPT Pro launched. Expect OpenAI and Anthropic to respond.
- Chrome agent is coming this summer. Google confirmed Spark will run “directly within Chrome” later in summer 2026, acting as an agentic browser. That puts it head-to-head with OpenAI’s ChatGPT-in-Chromium effort and Anthropic’s Claude Computer Use.
- Sub-agents and email-as-API. The roadmap to spawn custom sub-agents and accept work via email is a meaningful primitive for power users and small teams.
How to use Gemini Spark today
Spark is in trusted-tester beta this week and expands to US AI Ultra subscribers next week. If you have AI Ultra ($200/month after the cut) or AI Pro+ ($100/month new tier), here is the activation path.
- Open the Gemini app on Android, iOS, or gemini.google.com on the web. Sign in with the Workspace or personal account that holds your AI Pro+ or AI Ultra subscription.
- Navigate to Settings → Labs and Previews → Gemini Spark. Enable the toggle. You will receive your dedicated Spark email address (format:
yourhandle.spark@gemini.google.com) and a one-time Workspace permission consent for Gmail, Docs, Drive, Calendar, and Sheets. - Connect partner apps. Tap Connections and authorize Canva, OpenTable, Instacart, and any others Google enables in beta. Each connection uses standard OAuth and shows the exact scopes requested.
- Send Spark a first task by email — the simplest entry point:
To: yourhandle.spark@gemini.google.com
Subject: Q2 status doc
Pull every email from gmail tagged "Q2 status" since April 1
plus the meeting notes in Drive tagged "Q2 review" and create
a single Google Doc summarizing decisions, owners, and open
risks. Reply to me with the doc link when done.- Spark replies within minutes with a draft summary of what it will do and asks for confirmation if any side effects are detected (email sends, purchases, calendar invites, file shares). You confirm by replying “yes” or by tapping a confirmation card in the Gemini app.
- For multi-step automation, use the Spark Workflows panel in the Gemini app. The pattern below schedules a recurring weekly digest:
Workflow name: Weekly inbox triage
Schedule: Every Monday at 8:00 AM local time
Trigger: schedule
Steps:
1. Scan Gmail for unread emails from the past 7 days
2. Group by sender importance (use Spark contact-rank)
3. Draft a Doc summarizing must-reply, FYI, and noise
4. Email me the Doc link
5. Do NOT auto-reply to any thread
Confirmation required: false (read-only workflow)- To revoke Spark access at any time, go to Settings → Connections → Manage Spark. Revocation is immediate and ends all in-flight sub-agent tasks.
How Gemini Spark compares to the agent field
| Capability | Gemini Spark | ChatGPT Agent Mode | Claude Computer Use | Microsoft Copilot Agents |
|---|---|---|---|---|
| Base model | Gemini 3.5 | GPT-5.5 Instant | Claude Opus 4.7 | GPT-5.5 / Claude Opus 4.7 |
| Native Workspace access | Yes — Gmail, Docs, Drive, Calendar, Sheets, Slides | Via plugins | Via MCP servers | Native to Microsoft 365 only |
| Persistent cloud execution | Yes — runs without local device | Limited (sessions) | Local container required | Yes within Microsoft 365 |
| Email-as-API | Yes — dedicated Spark address | No | No | No |
| Browser agent | Chrome integration summer 2026 | Atlas browser shipped | Computer Use API | Edge Copilot |
| Sub-agent spawning | Roadmap | Limited | Via Claude Agent SDK | Agent 365 supports |
| Confirmation gates | Mandatory for spend/send | Configurable | Per-action prompts | Tenant policy controlled |
| Entry price | $100/month AI Pro+ | $20-$200/month | API metered + Claude.ai $20 | $30/user Microsoft 365 Copilot |
| Top tier | $200/month AI Ultra | $200/month ChatGPT Pro | $200/month Claude Max | $30/user + Agent 365 $15 |
Spark’s clearest advantage is native Workspace integration with no setup. Its clearest gap is the absence of a shipping browser agent — Atlas, ChatGPT’s browser, is already in users’ hands, and Anthropic’s Computer Use API is in production at multiple enterprises. The Chrome integration timeline matters; if Spark slips past summer, the agentic-browser race shifts decisively away from Google.
What’s next for Gemini Spark and Google’s agent stack
Three things to watch over the next 90 days. First, the trusted-tester expansion: Google said Spark goes to AI Ultra subscribers next week and to AI Pro+ in the following weeks. The pace of that ramp will tell you whether the agent is production-ready or whether it is being throttled while reliability catches up. Second, the partner-app list. Today’s beta connects Canva, OpenTable, and Instacart. A real agent platform needs hundreds of integrations; expect rapid additions through Q3 and pressure on Google to publish an open connector standard (likely MCP-compatible given Anthropic’s protocol momentum).
Third, Chrome. The promise that Spark will run inside Chrome later this summer is the load-bearing claim. ChatGPT’s Atlas browser shipped in Q1 2026 with a small but vocal user base. If Google ships a Chrome-Spark integration that converts the existing 3 billion+ Chrome users into agent users by default, the consumer agent battle is effectively over. If the summer date slips, Google cedes the surface that defines the category.
Underneath all of this, Gemini 3.5 — the model powering Spark — is also the model Google is positioning as its mid-tier API offering. Gemini 3.5 Flash launched alongside Spark at I/O 2026 with what Google claims is one-third the price of comparable frontier models. The Flash tier is the part developers should pay attention to first; Spark is the consumer face of the same intelligence.
Frequently Asked Questions
Is Gemini Spark available outside the US?
At launch, no. Beta access through the next several weeks is US-only for AI Ultra and AI Pro+ subscribers. Google has not committed to an international rollout date but confirmed at I/O that English-language markets (UK, Canada, Australia) are the next expansion targets, followed by European and Asian markets through Q3 2026. Spark’s underlying Gemini 3.5 model is available globally via the Gemini app and API; only the agent layer is geographically restricted at launch.
What can Gemini Spark do that the existing Gemini chat cannot?
The core difference is persistence and cross-app action. Gemini chat answers questions and drafts content during a single session. Spark runs in the background, accepts work via email or workflows, executes across Gmail, Docs, Drive, Calendar, and connected partner apps, and confirms before taking any action with side effects. Spark also has memory across sessions and can spawn sub-agents for parallel work — features the chat product does not have.
How does Spark handle privacy and data access?
Spark uses standard Google Workspace OAuth scopes and only accesses the apps you explicitly connect. Each action with side effects (sending email, making purchases, sharing files) requires confirmation by default. Google says Spark conversations are not used for model training in the AI Pro+ and AI Ultra tiers. Enterprise Workspace customers can additionally configure Spark via the admin console with tenant-wide policy controls. You can revoke Spark access at any time from Settings, which ends all in-flight tasks immediately.
How does Gemini Spark compare to Microsoft Agent 365?
Microsoft Agent 365 — which went GA at $15/user earlier in May 2026 — is positioned for enterprise governance across Microsoft 365 with strong policy, audit, and identity primitives. Spark is positioned for consumer and prosumer use with native Google Workspace access and partner-app reach. Agent 365 wins where IT control matters; Spark wins where individual workflow integration matters. For a small team on Workspace, Spark is the obvious pick; for an enterprise on Microsoft 365, Agent 365 is.
Will Gemini Spark replace Google Assistant?
Eventually, yes. Android Central and several Google product leads at I/O confirmed that Spark is the architecture Google plans to graft onto the Assistant surface — phone, smart speaker, Android Auto, smart displays — over the next two to four quarters. Today’s Spark is the agent layer; bringing it to voice surfaces requires latency work and turn-taking refinement that Google has not yet shown publicly. Expect Assistant-to-Spark migration to start with Pixel devices late in 2026.
Can developers build on Gemini Spark?
Not directly today. Spark itself is a consumer product without a public SDK at launch. However, the Gemini 3.5 model powering Spark is available via the Gemini API and Vertex AI, and Google said at I/O that the agent runtime primitives (sub-agent spawning, persistent memory, tool calling against Workspace) will reach Vertex AI Agent Builder later in 2026. Developers who want agent capabilities today should use Vertex AI Agent Builder with Gemini 3.5 plus the Workspace APIs.