Google's Gemini Agent Gets Its Own Email, Its Own Drive, and Claude Underneath. Not One Price or Date Is on a Google Page.

Google Cloud CEO Thomas Kurian opened Gemini at Work 2026 this morning, October 8, 2026, with a product that is not a model: "To meet this moment, today we are excited to announce the Gemini agent: your new single, universal agent for work." His shorter version, to CIO Dive: "Today, Gemini becomes an agent. You give it objectives, not just instructions." One agent that answers questions, does knowledge work, makes media, and writes and runs code, "all in a single agent and a single API." It plans a job, spins up temporary sub-agents, and, if you ask it to, takes a permanent seat on the team with its own email address, calendar, Drive, and entry in the company directory.

The part we did not expect is the engine room. Google says the agent runs each job "on the model that fits best," and the roster today is the Gemini family plus Claude from Anthropic. OpenAI's Dots, the nearest sibling, run on GPT-6 Astra and nothing else. Google is shipping a rival's model as a peer inside its own agent, with its own spend caps metering the bill. That deserves a closer look than the keynote gave it.

The catch is simpler: nothing about when you can have it or what it costs is on a Google page. Private preview, no additional charge, and an autumn rollout are all Google statements relayed by reporters. We read Kurian's post, the Keyword summary, the Gemini Enterprise release notes and docs, the Vertex pricing page, and Artificial Analysis. Here is what held up.

One Agent, Hired Like an Employee

It keeps working after you log off. Google's post says the agent "runs in the cloud," keeps "a single set of memories, context, and one personalization graph" across devices, and that long work "keeps running after you close your laptop." It lists four memory types: session, semantic, procedural (how a job gets done, including skills the agent writes for itself), and episodic (everything it has done before). Those memories live in Google's layer, which matters later.

Sub-agents are disposable; coworker agents are staff. The blog says the agent "can dynamically create a roster of sub-agents," described as "temporary, job-specific agents, each with their own identity," coordinated through "parallel and sequential steps that can run for hours or days." A coworker agent is "more like a team member, with a persistent, defined role and operational presence across days, sessions, and multiple changing responsibilities." It gets "its own Workspace account, including an email address, calendar, Drive, and presence in your company directory," and coworker agents "only have access to the context that you or your team members provide." On stage, Kurian added two things the blog leaves out: "Each co-worker has its own email address, its own Drive, its own virtual machine, and its own browser." (Transcription by Cloud Wars.)

Two things Google did not say. The blog never calls the agent's mailbox Gmail (several headlines do); it says "email address" and "Workspace account." And it does not say how many coworker agents an organization can run, or whether a coworker with its own account consumes a license. VentureBeat asked; Google has not clarified.

Security is identity-shaped. Every agent gets "its own identity, cryptographically attested and governed like an employee, with least-privilege permissions," stamped into its logs and into any VM it spins up. Tasks run in an Agent Sandbox with its own network boundary, and traffic passes through "Agent Gateway, an AI network firewall enforcing your organization's policies in real time." It can also "connect and work securely with any Model Context Protocol (MCP) server inside or outside your company network," and teams can publish tools and skills to company-wide registries.

The Model Underneath Is a Separate Choice, and One of Them Is Claude

Kurian's line is the one to remember: "Gemini is the agent, and the model underneath it is a separate choice." The blog names the Gemini family by role, Argon for frontier reasoning, Flash for speed and volume, Omni for generative media, Gemma for open-weights edge work, "and Claude models from Anthropic today," with "other leading private and open models in the future." Smart routing picks per task: "Our routing tool automatically triages enterprise workloads so each one runs on the model that delivers maximum performance at the lowest possible cost." Google publishes no rule for how the Auto setting chooses. TechCrunch reports you can override it.

Which Claude. The only Google document that names versions today is the Gemini Enterprise release note entry dated October 8: administrators can enable Claude Opus 5.5 (claude-opus-5-5) and Claude Sonnet 5.5 (claude-sonnet-5-5) for AI developer tools. A model-menu list circulating on X today names Claude Opus 5 instead; it has no screenshot and no Google source. We print 5.5 and 5.5. Both are already on Vertex AI at Anthropic's list prices: Opus 5.5 at $4 in / $20 out per 1M since September 22, Sonnet 5.5 at $2 / $10 since September 28, and Haiku 5.5 at $0.10 / $0.50 (up to 100K tokens) since yesterday.

Which GPT. None. Kurian's keynote line, as transcribed by Cloud Wars, was broader: "The agent is separate from the model underneath, which is a choice: Gemini, Claude, GPT, or open models like GLM, whichever fits the job." The blog adapted from the same keynote says Gemini and Claude "today." Constellation Research's exact sentence is "Today, Gemini agent can route to the Gemini family of models and Anthropic's Claude, but Google Cloud said other models, closed and open will be available in the future." No Google document lists a GPT or GLM option. Treat the stage line as roadmap.

Why a rival's model. Because, on the numbers, the rival is ahead; see the next section. Anthropic has published nothing about the arrangement as of today. Nick Patience of The Futurum Group gave CIO Dive the caveat that outlives the press release: "CIOs should go in with their eyes open. Model choice is open, but the memory, context and skills an agent builds up over time live in Google's layer, and that's where the real switching costs will sit."

The comparison with OpenAI's Dots is the useful one. Dots, launched at DevDay on September 29, are always-on agents that run on GPT-6 Astra; one is included with the ChatGPT Pro tiers and Business Premium, extra dots are promised for later with no price, and the model is not a choice. Google's agent is model-agnostic by design and, by Google's statements to the press, has no SKU of its own. Neither company has published a price list for the agent itself.

The Benchmarks, Sorted by Who Reported Them

There is no benchmark for the Gemini agent. Google published no task-completion rate, no eval, and no system card for it. What exists are independent scores for the models the router can send your work to, and Google's own scale numbers.

Independent first. Artificial Analysis's Intelligence Index is still v4.3.2, the version in place since September 19, so these scores are comparable with our September coverage. The picker's rows, as AA lists them:

Three readings. First, the top of Google's own model picker belongs to Anthropic: Opus 5.5 leads the index at 58, Sonnet 5.5 is second at 56, and Google's best, Argon, ties GPT-6 Astra and Claude Fable 5.1 at 53. Second, Argon is the value row: AA's September 30 line is that it "equals GPT-6 Astra on the Artificial Analysis Intelligence Index at 60% of the Cost per Task," $1.99 against Astra's $3.26, with a 15% hallucination rate AA calls "the lowest of any model scoring 45+ on the Intelligence Index." AA also measured 78% on AutomationBench-AA, 57% on Terminal-Bench 4.0, and 1494 Elo on AA-Briefcase. Third, the tie carries an asterisk: the Claude rows run at max effort with Anthropic's default fallback enabled, Argon runs at high with no fallback, and the $1.99 rests on a launch discount AA describes as lasting "at least one month."

And the row that is not for sale. AA's Argon page, as of today, says the model is "currently being rolled out to selected users and is not publicly available." Google announced it on September 30 for Fairwind cyber defenders first. The Google model you can actually put on a purchase order today is Gemini 3.8 Flash at 41, or Gemini 3.1 Pro Preview at 30 on the same index. That is the plainest answer to why Claude is in the picker.

Now the vendor numbers. These are Google's self-reported figures from Kurian's post and keynote, none independently measured:

Three honest observations:

• None of it measures the agent. Token volumes, adoption rates, and agent counts describe usage of the products Google already sold. The one customer named as a tester of the new agent, On, appears in a press release with no outcome figure.
• The router is the product, and it is a black box. Google's cost argument rests on Auto choosing the cheapest adequate model. It has published nothing about how that choice is made, and its customer examples (On, Shopify, PayPal) describe adopting routing, not what it saved.
• The frontier Gemini is not generally available. Argon is named first in the family and is the only Gemini that reaches the Claude tier on AA. On launch day you cannot buy it.

The Spend Cap Has a Few Minutes of Give

The blog's wording is firm: "You can set a hard limit on a project's AI spend in the Cloud Billing Console, and Gemini enforces it." If the cap trips, "that project's agent pauses, and you can choose to resume work with a single click in the console." Per-project tracking means departments get charged back. Kurian, to The Register: "You set its budget before it starts and that budget covers the whole cost of the work." The Register notes the controls arrived about two months ago; Google's August 26 FinOps post describes a "firm monthly spend limit," email alerts at 50%, 80% and 100% of budget, a temporary pause of the agent's API calls at the limit, and optional overages drawn against a Flexible Savings Plan (10% off for a one-year commitment, 20% for three).

Google's own Gemini Enterprise doc, "Manage costs," softens two words. The cap is a project-level budget scoped to Vertex AI and covers the Gemini Enterprise app, the Agent Platform, and AI coding tools. Then: "Because stopping usage can take a few minutes to take effect, you might incur charges that exceed your limit." And: "Feature access resumes automatically when the quota resets or at the start of the next billing cycle," where the blog describes a manual one-click resume. And the gate: "Overages and spend controls are available only if your project has an invoiced Cloud Billing account and at least one active, non-free-trial subscription." So the hard cap is a soft cap with a lag, and card-on-file accounts do not get it.

One more scope note. The cap governs token and sandbox spend. It says nothing about money an agent with its own identity might move inside Salesforce, ServiceNow, or a payments system it is connected to. That is a permissions problem, not a billing one, and Google's answer to it is the least-privilege identity above.

The Pricing Fine Print, Such As It Is

No Google page states a price, an entitlement, or a SKU for the Gemini agent. Kurian's post contains no availability sentence at all; we searched it for "preview," "available today," "coming soon," and "no additional charge." Only the industry editions carry "now in preview."

What exists is a statement Google confirmed to VentureBeat: the agent "will be included at no additional charge for customers where Gemini Enterprise is available, including through existing Google Workspace business and enterprise subscriptions." It "will not require a separate application, website or setup process," and Google "will share additional pricing details closer to the wider release." Constellation Research heard the same shape: "Consumption based pricing will be the norm for Gemini agent," and "Google Cloud isn't doing an extra SKU or add-on." 9to5Google adds that wide availability is coming "for Workspace customers with select Business and Enterprise plans," and Google has not listed which plans qualify.

Read together: the seat comes with Gemini Enterprise or a qualifying Workspace plan, the tokens and sandbox time are metered under the caps above, and Claude usage on Vertex is pay-as-you-go at Anthropic's list prices. What VentureBeat could not get answered: whether coworker agents with their own accounts carry licensing fees, whether long autonomous tasks and sub-agents bill differently, and how many coworkers you can deploy.

The seat price itself is a dash. Google's Gemini Enterprise editions page lists Business, Standard, Plus, Pay-as-you-go, and Frontline, with pooled storage figures and no dollar amount anywhere. The per-user figures in circulation are Google Cloud partner price lists, not Google's page, so we will not print them as Google's.

Where You Can Run It (and Where You Cannot, Yet)

CIO Dive's sentence, from a Google email: "The Gemini agent is available today in preview to a select set of customers, the company said in an email, with wider availability coming soon." A Google Cloud spokesperson told VentureBeat the same, "with wider availability promised soon but no specific date given." Named testers: On (Google's press release calls it "a trusted tester of the new Gemini agent"), Shopify, and PayPal. Constellation's own estimate: "General availability for Gemini agent, currently in private beta, is expected around the end of October or early November." Decrypt's summary is the accurate one: Google published no pricing and no general availability date. Update (Oct 9): Futurum reports that Google expects general availability by the end of October 2026, starting in North America; that date still appears on no Google page.

Consumers wait. TechCrunch: Google "will initially focus on bringing the agent to businesses before later rolling it out to consumers," which Pichai framed as solving the "harder problems around security, scale, and performance" first. The consumer Gemini app already has a separate agent, Gemini Spark (May 19, a "24/7 personal AI agent" on Gemini 3.5 and the Antigravity harness), and the Gemini Enterprise app already has Agent Designer, long-running agents, and an Inbox from Next 26 in April. Today's post mentions none of them. Patience's read, via CIO Dive: Google "talked about Gemini Enterprise as a 'front door' for AI a year ago, but what sat behind that door was a set of tools employees still had to assemble themselves," and "This release puts a single agent behind the door that does the assembly."

Once it is wider, the surfaces are the long list in the table, plus Antigravity for developers. One nuance: ServiceNow is in Kurian's spoken list of surfaces and in TechCrunch's, but in Google's written post it is a connected system "like Salesforce and ServiceNow," not a channel. The industry editions are further along: Gemini Enterprise for Financial Services ("more than 50 foundational skills"; data from FactSet, LSEG, S&P Global, and SEC filings; CME Group and Deutsche Bank as early users) and for Legal (matter-level permissions and ethical walls inherited from NetDocuments and iManage; Harvey, Onit, and Cooley building on it) are "now in preview," with Government, Healthcare, and Retail "coming soon."

EU availability and data residency for the agent, its sandboxes, and coworker storage: dash. No Google page addresses them.

Same-day housekeeping on the Gemini API: gemini-3.7-flash is deprecated and routes to gemini-3.8-flash, gemini-3.5-flash routes to gemini-3.6-flash, and the deep-research-pro-preview-12-2025 agent shuts down October 23.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use it if you are already a Gemini Enterprise or Workspace Enterprise shop, you can get on the preview, and your workloads are the kind a router can grade: document pipelines, reporting, ticket and inbox triage across Workspace, Microsoft 365, and Slack. Set the per-project cap on an invoiced account before the first coworker agent gets an email address, and budget for Claude at Anthropic's list prices, because the best model in the picker is not Google's.

Skip it for now if you need a date, a price, or a contract: none exists. Skip it for anything where a few minutes of overshoot on a "hard" cap matters, for EU data-residency cases Google has not addressed, and for any plan that depends on Argon, which you cannot buy today. If your agents must stay on one vendor's models, Dots is the single-vendor answer and has the same missing price list.

The frontier is elsewhere, and also in the picker: Claude Opus 5.5 (max) leads the v4.3.2 index at 58 with Sonnet 5.5 at 56; Google's Argon ties Astra and Fable 5.1 at 53 and is not generally available. Our model comparison has the current standings side by side.

Update, October 15, 2026

• October 9: The Futurum Group reports a Google representative said the company "does not intend to charge a per-seat license when the Gemini agent acts as a coworker in Workspace." VentureBeat (October 8) and Creative Strategies (October 12) say Google has not clarified. Nothing on a Google page.
• October 12: Memeburn reported that the launch video's model menu shows Claude Opus 5, not 5.5; OrcaRouter (October 9) had already traced that list to a single @testingcatalog post on X. Google's release note three days later settles it below.
• October 13: Gemini 4 Argon went generally available as gemini-4-argon on the Gemini API, AI Studio, Antigravity, and Gemini for business on web, desktop, and mobile, at $2 in / $10 out per 1M through November 30, 2026, then $4 / $20 from December 1 (cache $0.10 then $0.20; Batch half), which puts it level with Claude Opus 5.5 on Vertex after the step. Specs: 1M unified input-plus-output context, March 2026 knowledge cutoff, thinking levels low, medium (default), and high; Vertex retirement "October 13, 2027 or later." The Gemini Enterprise release note says the Argon toggle "defaults to on beginning October 27, 2026." The same day, the Vals Index v2.1 placed Argon first at 68.90%, ahead of Sonnet 5.5 (67.04%) and Opus 5.5 (66.97%), and the consumer Gemini app gained a background "tasks" mode. Google Cloud Japan told ITmedia the Japan rollout timing is undecided.
• October 15: The Gemini Enterprise release note says administrators can now make Claude Opus 5.5 and Claude Sonnet 5.5 available to users in the Gemini Enterprise app's model selector, "charged on a pay-as-you-go basis," in the global location and the us and eu multi-regions; Google's quotas page prices them in the app at $4 / $20 and $2 / $10 per 1M, available "only with pay-as-you-go usage." The same day, Google Cloud's startups post said the agent's capabilities "are not publicly available yet." Through October 16, the Gemini Enterprise release notes contain no Gemini agent entry, no date, and no price.

Sources

• Google Cloud: Welcome to Gemini at Work 2026, introducing the Gemini agent (Thomas Kurian)
• Google Cloud: Gemini Enterprise release notes
• Google Cloud: Gemini Enterprise, manage costs
• Google Cloud: Gemini Enterprise editions
• Google Cloud: Vertex AI generative AI pricing (Claude list prices)
• Artificial Analysis: Gemini 4 Argon
• VentureBeat: Google Cloud unveils persistent Gemini agents (Google-confirmed pricing statement)
• CIO Dive: Google's Gemini agent, Kurian and Nick Patience
• Cloud Wars: Kurian keynote transcription

*Weighing an agent that picks its own model against one locked to a single vendor? Take the free 2-minute quiz and get matched. Then read the sibling that launched nine days earlier with a missing price list of its own: OpenAI Dots.*