Google adds a "Task" mode to the Gemini app that can make calls, pay bills, shop and manage schedules, using a new Gemini 4 Hydrogen model
Edited by Elliot Marsh, Editor · written with AI from the linked sources (how)
Google is rolling out agentic task handling in the Gemini app. Users switch from Chat to Task, text Gemini on Android or iOS (RCS), or forward emails to gemini@google.com. Gemini can call businesses (disclosing it is an AI; businesses can opt out), book appointments on Google Calendar, analyze utility bills via a Google Password Manager integration, and shop and check out through Google Pay using the Agent Payments Protocol (AP2). It will work with Mastercard and Visa, with American Express coming soon, and users can set a spending cap. Receipts go to Google Wallet, which is expanding to iOS. Gemini connects to 60+ apps, and a Connected Apps Platform is open to developer applications. Tasks run in per-user hardware-isolated VMs on Google Cloud's Agent Compute Engine, with prompt-injection filters, and Gemini pauses for confirmation before critical steps such as "Buy". The basic update is free in the U.S., Japan, Korea, Canada, Australia and India on the new Gemini 4 Hydrogen model. Advanced features start in the U.S. The Universal Commerce Protocol (UCP) is planned next.
Entities: Google, Gemini, Gemini 4 Hydrogen, Josh Woodward, Agent Payments Protocol (AP2), Universal Commerce Protocol (UCP)
1 primary
What happened
Google is adding a "Task" mode to the Gemini app, running on a new model it calls Gemini 4 Hydrogen. Gemini can phone businesses (it says it is an AI, and businesses can opt out), book appointments on Google Calendar, analyse utility bills through Google Password Manager, and shop and check out via Google Pay using the Agent Payments Protocol (AP2). Visa and Mastercard are supported, American Express is "coming soon", and users can set a spending cap. The basic update is free in the U.S., Japan, Korea, Canada, Australia and India, advanced features start in the U.S. only, and the Universal Commerce Protocol is planned later. The source is Google's own blog, and no independent testing is cited.
Why it matters
This puts agents that act, spend money and phone third parties into a mass-market assistant, not a niche tool. Google's advantage is that it already holds users' email, calendar, passwords and cards, so it needs fewer integrations and permissions than rivals. Retailers, small businesses that take phone bookings, and payment firms now have to decide how to handle AI callers and AI buyers, and the Connected Apps Platform gives developers a route in. The real effect depends on how often tasks finish correctly, and nothing public yet shows that.
What is noise
The claim that this saves time and money for "hundreds of millions of people" is an aspiration, not a measured result, and "best assistant" is marketing. Security assurances (isolated VMs, prompt-injection filters, confirmation before "Buy") are Google's own description and unverified. Missing context includes failure rates, how liability works when an agent makes a wrong purchase or booking, and how many people will actually get the advanced features outside the U.S.
Watch next
- 01Independent hands-on tests of Task mode: completion rates, wrong purchases or bookings, and how well the spending cap and confirmation prompts hold up against prompt-injection attacks.
- 02Business response to AI phone calls: how many opt out, whether other platforms and retailers adopt AP2 or UCP, and when American Express and the UCP launch.
- 03Expansion beyond the U.S. for advanced features, plus any usage figures from Google, and competitor moves on agentic payments from OpenAI, Apple or Amazon in the next quarter.
Coverage
2 storiesNewsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Mistral releases Mistral Large 4, a ~1T-parameter open-weight multimodal MoE model, with day-0 Hugging Face support29 Oct 202680
- Meta releases Muse Spark Open, an open-weight ~816B-parameter multimodal mixture-of-experts model, with Hugging Face support15 Oct 202680
- OpenAI discloses sandbox-escape and credential-leak incidents, confirms pause on tool-use for its most capable models26 Sept 202680