Gemini 4 Argon Reaches Pro and Ultra Subscribers as Google Adds a Tasks Mode [1]

Google brought Gemini 4 Argon to Gemini app Pro and Ultra subscribers and introduced a mode that continues working after a user steps away. That gives companies something to evaluate, but the report does not establish whether tasks mode includes the permissions, supervision or administrative controls a business workflow may require. ¹

What changed in Gemini Pro and Ultra

Google brought Argon to Pro and Ultra subscribers on October 13, 2026, two weeks after it first went to vetted cybersecurity defenders. Pro access was still rolling out when the report was published, and Google’s plans page said tasks mode was limited to select countries. ¹

In the U.S., Google AI Pro costs $19.99 a month; Ultra plans start at $99.99 and include higher Argon limits. The report does not specify those limits or say whether Argon and tasks mode work through business-managed Google accounts. ¹

That gap matters. A consumer subscription lets an individual try the feature; it does not establish whether an organization can provision, govern or audit it centrally. Until those details are clear, treat this as a subscriber rollout, not an enterprise launch. ¹

What Gemini tasks mode actually does

Pro and Ultra subscribers can switch tasks mode on with a toggle in the main chat window. Google says Gemini can keep working on a job after the user steps away from the computer. ¹

The examples go beyond drafting text. Gemini can place a grocery order through Instacart or book a table through OpenTable. The report says the feature connects to more than 40 apps, including Adobe and Shopify. ¹

Tasks mode can therefore reach connected-app actions, not just generate chat responses. The report does not specify how long a task can run, how many can run concurrently, what happens when a step fails, or how much a task can finish. ¹

It also does not say whether Gemini asks for confirmation before a purchase, booking or other consequential action. The examples show possible actions, but do not describe the safeguards around them. ¹

For a business team, “continues a job” is not an operating model. Before using the feature in a workflow, teams need to know what the agent can access, when it must ask first, and how a person can intervene if something goes wrong. Our piece on where AI agents’ access boundaries should sit covers that question in more depth.

How Gemini 4 Argon compares on the DeepSWE benchmark

Google reported that Argon scored 77.9% on DeepSWE v1.1, described as a long-horizon software engineering test, compared with 74.2% for Claude Opus 5.5. ¹

Model	Reported DeepSWE v1.1 score
Gemini 4 Argon	77.9% 1
Claude Opus 5.5	74.2% 1

That is a signal about performance on the reported coding benchmark. It does not, by itself, establish how safely a model handles connected-app actions. ¹

The report describes DeepSWE v1.1 as a long-horizon software engineering test. It provides no evidence that the benchmark evaluates purchase or booking confirmations, connected-app permissions, task controls, administrative review, data handling or recovery from failure. Those operational details remain unspecified in the report. ¹

Google also said Argon gives more direct answers with less hedging, repeats itself only when needed, and provides progress updates more often on longer jobs. These are product claims attributed to Google in the report, not independent findings presented there. ¹

A higher coding score can justify testing Argon on software tasks. It does not establish what the task runner is allowed to do, who can supervise it, or what happens when something breaks.

Supervision and access questions still open for tasks mode

The report describes a toggle for starting tasks mode, but does not say whether users can pause, cancel or resume a task after it starts. ¹

It also does not describe what permissions users grant to connected apps, whether an administrator can restrict access by app, action or data type, or what task activity is logged for later review. ¹

These are practical questions for teams assessing the feature. If a team cannot establish whether a task is reversible, reviewable or limited to approved data and actions, it cannot set a clear boundary for sensitive work. Treat the following as questions to resolve before deployment, and see runtime governance controls for always-on agents for what such controls can look like:

• Connected-app permissions, including whether access can be scoped by app, action or data type ¹
• Confirmation requirements before purchases, bookings and other consequential actions ¹
• Whether users can monitor, pause, cancel or resume a task, and how failures are handled ¹
• Task duration and concurrency limits, and what activity administrators can audit ¹
• Business-account support and the countries where tasks mode is available ¹

Does Gemini tasks mode fit company workflows yet?

For individual evaluation, the case is reasonable: Argon’s reported coding score is a signal on a long-horizon software engineering test, and the Instacart and OpenTable examples show tasks mode reaching connected-app actions. ¹

For production use, the report leaves important questions unanswered. Pro access was still rolling out, tasks mode was limited to select countries, and the report did not confirm business-account support. ¹

Keep model quality and operational readiness separate. A benchmark score does not say who can authorize an action, whether a human reviews it first, or how an administrator investigates a failed task. Until those answers are documented, avoid routing purchases, bookings or other commitments through tasks mode. If you are unsure an agent is the right tool at all, when to skip an AI agent is a useful starting point.

What to do before piloting AI task automation

Start by checking whether tasks mode is available in your users’ countries and whether Argon has reached their accounts. The report says the feature is limited to select countries and that Pro access is still rolling out, but does not provide a country list or a completion date. ¹

Test in stages: begin with non-consequential tasks, then assess connected-app permissions and confirmation behavior before allowing the system to touch records, place orders or create commitments.

Ask Google for documentation on task duration, concurrency, pause and cancellation options, error handling, connected-app scope, activity logs and administrative controls. The report does not answer those questions. ¹

Evaluate Antigravity separately. The report says Google’s agentic coding interface runs inside the Gemini app for macOS and does not mention a Windows version. That is a separate consideration for teams assessing coding tools, not evidence about how tasks mode is governed. ¹

FAQ

When did Gemini 4 Argon reach Pro and Ultra subscribers?

Google brought Argon to Gemini app Pro and Ultra subscribers on October 13, 2026, after an earlier release to vetted cybersecurity defenders. Pro access was still rolling out when the report was published. ¹

What is Gemini tasks mode?

It is a mode available to Pro and Ultra subscribers that can be turned on from the main chat window and is described as continuing a job after the user steps away. The report gives grocery ordering through Instacart and table booking through OpenTable as examples. ¹

Is Gemini tasks mode available everywhere?

No. Google’s plans page said tasks mode was available only in select countries, and the report does not name them. ¹

Does Gemini tasks mode ask for approval before purchases or bookings?

The report does not say whether Gemini requests confirmation before purchases, bookings or other consequential actions. ¹

Is Gemini tasks mode ready for company workflows?

The report does not establish business-account availability, administrator controls, permission restrictions, monitoring, cancellation behavior or failure handling. Treat those as unresolved evaluation requirements rather than inferring readiness from Argon’s benchmark score. ¹

Watch for Google to publish the country list, confirm business-account support, and document task limits and audit controls. Until then, the benchmark speaks to coding performance, not deployment risk. ¹