A hero title card for the Gemini Spark Computer Use leak report, with a LEAK REPORT badge, the subtitle Desktop Control, Drive Backup, What's Still Unverified, three chips reading Unreleased, Unverified: settings-trace leak, and Computer Use · Gemini Spark, and two cards: a Reported card noting Computer Use spotted in Gemini Desktop testing on August 18, and a Confirmed card noting Gemini Spark is live and now runs on Gemini 3.7 Flash.
Guides & Insights

Gemini Spark Computer Use Leak: Desktop Control, Drive Backup, and What's Still Unverified

Author

Alistair Wren

Date Published

Latest models · 20View all models
Benchmarks: Artificial Analysis · updated daily
Back to all posts

The most interesting thing about the newly leaked Computer Use mode for the Ge​mini Desktop app is not that Ge​mini Spark will be able to drive other applications on your machine — it is the guardrail Go​ogle appears to be building to make that safe. TestingCatalog, the outlet that first surfaced Spark's desktop ambitions in May, says it found traces of a Computer Use option in the desktop app's settings on August 18, apparently in testing with a limited group of trusted testers. None of that is confirmed: Go​ogle has said nothing, and the entire feature set rests on one outlet's reading of an unreleased build. What is confirmed is the agent behind it — Ge​mini Spark, Go​ogle's 24/7 personal assistant, is live for AI Pro and Ultra subscribers and has just moved onto the freshly released Ge​mini 3.7 Flash. The leak splits cleanly in two: a reported feature, still unverified, and a shipped agent, already here.

What the leak says — and what it doesn't

TestingCatalog's August 18 report is a settings trace, not a screenshot of the feature working. The site says it found evidence of Computer Use in the Ge​mini Desktop app's configuration and speculates Go​ogle may be running it with a limited set of trusted testers. No date, no launch window, no confirmation from Go​ogle, and no independent second source — read every detail below as reported-but-unverified, not as fact:

• Scope — Computer Use would let Ge​mini Spark control other applications and access files in selected local folders, the same class of capability rival desktop assistants have been rolling out this year.

• Where it lives — reported to appear as a "Task" option in the prompt-bar selector (next to the existing image and video generation controls), not as a navigation-bar toggle. This matches Spark's existing Tasks/Skills/Schedules mental model.

• Oversight — users would be able to watch the agent's actions in a mini-view of the active screen, keep the machine from sleeping while a task runs, and trigger tasks from another device.

• The safety layer — an "Advanced Backup" option that automatically backs up the folders Ge​mini Spark will touch to Go​ogle Drive before it modifies anything, with a companion setting to stop and ask for approval if the backup cannot complete.

A single-column scoreboard titled Gemini Spark Computer Use — the scoreboard. Rows read Status: in testing, not shipped; Evidence: settings-trace leak, 2026-08-18; Scope: drive apps, read selected folders; Guardrail: Advanced Backup to Google Drive; Underlying model: Gemini 3.7 Flash (live Aug 13); Release date: none announced. Footer reads Feature leak unverified; Gemini 3.7 Flash release vendor-confirmed.

The scoreboard above is the honest state of the evidence: six rows, and five of them are about a feature that has not shipped. The one row that is a shipped fact is the model underneath. Treating the feature as imminent would be reading too much into a settings panel — Go​ogle has leaked features that never shipped, and TestingCatalog's own coverage of the desktop app has been careful to separate what is wired up from what is actually live.

The guardrail is the story: an undo button, on by default

The "Advanced Backup" detail deserves more weight than the computer-use claim itself, because it is the rare leak about how a lab is trying to make an agent safe — not just what the agent can do. As reported, Ge​mini Spark would copy any file it is about to modify in a connected folder to Go​ogle Drive before touching it, and if it cannot complete that backup, it stops and asks for approval. In effect, every file edit gets a snapshot you can restore, switched on by default. That is the kind of guardrail users had to bolt onto other desktop agents themselves.

It also answers a real fear. The summer of 2026 has been full of agentic tools that silently reorganized or overwrote files with no recovery path, and the backlash landed on every lab shipping one. A backup-before-touch rule is a direct, cheap answer to the "it deleted my files" class of failure — and it is consistent with something Go​ogle has already shipped: the macOS Spark app launched in June with an "Alert when a backup fails" setting enabled by default and a "Keep this Mac awake to run tasks" control, and Spark is explicitly flagged as experimental and may make purchases without confirmation in some workflows.

The catch is practical. Backing up a selected folder to Drive before a task begins is only fast when the folder is small. A directory with hundreds of gigabytes would mean a long upload and a real chunk of storage quota consumed before the agent has done anything — which is exactly the trade-off TestingCatalog flags. The guardrail is a good default; the default's cost scales with your folder size.

Computer use is the next step, not the first one

The reason the leak is plausible — even though it is unverified — is that Ge​mini Spark has been moving toward exactly this since launch. Go​ogle previewed Spark at I/O on May 19-20 as a 24/7 agent that runs in the cloud on Go​ogle's machines and acts across Workspace. What followed is a sequence of steps that progressively bring the agent onto the user's own device:

• June 30 — Spark arrived in the Ge​mini app for macOS (US beta, Apple Silicon, macOS Sequoia+), where it could read and organize local files in connected folders, run scripts, and extract data into Sheets. Option+Space summons it; Command+Command grants it access to on-screen content.

• July 30-31 — Spark expanded to AI Pro subscribers in more than 160 countries, and Go​ogle announced Chrome integration.

• August 3 — "Chrome auto browse" began rolling out in the US, letting Spark drive the user's own Chrome browser using accounts they are already signed into and passwords saved in Chrome (with explicit permission), pausing before payments and other sensitive actions.

• August 13 — Ge​mini 3.7 Flash shipped and became the model Spark runs on, with Go​ogle positioning the Flash tier squarely at agentic, multi-step work.

Controlling the whole desktop, rather than one browser, is the logical extension of that arc — and the "Advanced Backup" guardrail is the answer to the question that arc raises: what happens when the agent can touch anything on your machine. Remote execution ("assign a task from your phone, Spark runs it on your Mac while you're away") was already flagged as coming soon in the macOS rollout, which is the same mechanism the leaked desktop app would need.

The model underneath: the Flash tier keeps getting cheaper

Ge​mini Spark is an agent, not an API model, but it runs on one: Ge​mini 3.7 Flash, released on August 13 at $0.75 per million input tokens and $3.75 per million output — exactly half of what Ge​mini 3.6 Flash cost at its own launch, on the same 1-million-token context window. Go​ogle's positioning for the Flash tier is unapologetically agentic: complex coding, multi-step execution, tool use, and reliability on long workflows. That is the workload Computer Use on a desktop is made of.

The performance picture needs the usual labeling. Go​ogle's own launch claims for Ge​mini 3.7 Flash are vendor-reported, and the first independent read came from Artificial Analysis, which scores Ge​mini 3.7 Flash at 56 on the Intelligence Index at high reasoning effort — four points above where Ge​mini 3.6 Flash landed at its own launch (52), and on the Pareto frontier of intelligence versus speed. Ge​mini 3.6 Flash launched in July at $1.50/$7.50 on the same 1M context window, which makes the 3.7 step both a capability gain and a 50% price cut on the tokens the agent actually spends. On the cost side, Go​ogle says the Flash line has been cutting token usage with each step — the 3.6 release alone was claimed to use roughly 17% fewer output tokens than Ge​mini 3.5 Flash on multi-step workflows. Those are vendor figures, worth reading as claims rather than facts.

A screenshot of the Artificial Analysis model page for Gemini 3.7 Flash, showing its Intelligence Index of 56, an input price of $0.75 per 1M tokens and output price of $3.75 per 1M tokens, a 1 million token context window, a speed of 372 tokens per second, and a release date of August 2026.

The capture above is the independent record against which the agent's underlying tier is measured today — Artificial Analysis's current read on Ge​mini 3.7 Flash, the model Ge​mini Spark now runs on. A leak can describe a feature; it cannot move a leaderboard. Until Computer Use actually ships, the measurable part of this story is still the model underneath it.

What's still unverified — and what to watch

Almost everything about the Computer Use feature is unverified. There is no Go​ogle confirmation, no release date, no confirmation that trusted testers are even running it, and no guarantee the current design survives to launch. What the leak does give you is a checklist of things to look for, because each one is externally visible before a formal announcement:

• A "Task" option appearing in the Ge​mini Desktop prompt-bar selector, next to image and video generation. That is the reported entry point.

• An "Advanced Backup" toggle in the desktop app's settings, or a change in how connected folders are labeled and backed up.

• The Windows desktop client gaining the same Computer Use surface — the macOS app shipped first last time, and Windows followed.

• The macOS Spark app's "Alert when a backup fails" setting turning into something larger, which would corroborate the Drive-backup design before any official launch.

• Geographic availability. Spark is currently not available in the European Economic Area, the UK, Switzerland, or Nigeria, and Computer Use will almost certainly launch inside the same regional gates.

The developer read: you can build this today with an API

The consumer feature is unreleased, but the underlying capability is not new at the developer level. Go​ogle already ships a Computer Use toolset in its Agent Development Kit that drives a Chromium browser through Playwright — screenshots, clicks, typing, navigation — on top of a Ge​mini model, and the Ge​mini API has had computer-use preview models since 2025. Desktop agents are not waiting on Go​ogle's consumer rollout; they are already being built on the model tier underneath it.

That is where a routing layer does the work the guardrail does for consumers. Computer-use and multi-step agentic workloads are exactly where model reliability is hardest to trust and where a single-model bet is riskiest — the agent equivalent of letting an unproven model reach into a folder with no undo. The engineering analog of Go​ogle's backup-by-default is automatic failover: try the capable-but-unproven model on a small slice of real traffic, and let a proven model catch the failures. On OrcaRouter, both sides of that setup are one API call away — Ge​mini 3.6 Flash and Ge​mini 3.5 Flash are live at Go​ogle's list price with zero markup, so a price change on the vendor side is reflected the same day, and the routing DSL lets you compose several models into a single call. When Ge​mini 3.7 Flash lands on the platform, moving to it is a configuration change rather than a re-integration. For a capability this new, the honest way to evaluate it is also the safe way: route a small slice of real work through it with failover, and let your own workload decide instead of a leak.

A screenshot of the OrcaRouter model page for google/gemini-3.6-flash showing the Vision, Tools, JSON and Reasoning capability chips, the /v1/chat/completions endpoint, an input price of $1.50 per 1M tokens and output price of $7.50 per 1M tokens, and a p50 time-to-first-token figure.

FAQ

Is Computer Use live in the Ge​mini Desktop app?

Not that anyone can confirm. TestingCatalog reports finding traces of the feature in the app's settings on August 18 and speculates Go​ogle may be testing it with a limited group of trusted testers. Go​ogle has made no announcement, and the feature should be treated as unverified.

When will Ge​mini Spark get Computer Use?

Nobody has said. There is no release date in the leak, and Go​ogle has not acknowledged the feature. The pattern to watch is the reported "Task" option in the desktop prompt bar and an "Advanced Backup" toggle in settings — those would appear before any official announcement.

What model does Ge​mini Spark run on?

Ge​mini 3.7 Flash, which Go​ogle released on August 13 — the version Spark now uses for AI Pro and Ultra subscribers, after the agent had previously run on Ge​mini 3.6 Flash and Ge​mini 3.5 Flash. Ge​mini 3.7 Flash is the model behind Spark's current agentic behavior, priced at $0.75/$3.75 per million tokens.

Where is Ge​mini Spark available?

In the US it is available to AI Pro and Ultra subscribers, and AI Pro access has expanded to more than 160 countries. It is not available in the European Economic Area, the UK, Switzerland, or Nigeria, and requires a personal Go​ogle account, 18+ age, and either Windows 11 or macOS depending on the client.

Bottom line: the Computer Use leak is real coverage, but it is coverage of an unreleased feature. Everything about it — the app control, the folder access, the Advanced Backup guardrail — is a settings trace from one outlet, and Go​ogle has confirmed none of it. What is confirmed is the agent and the model: Ge​mini Spark is live, and it now runs on Ge​mini 3.7 Flash, the cheapest Flash tier yet at half the price of its predecessor. For a reader choosing a model, that is the part you can act on. For a developer who wants the computer-use capability before Go​ogle's consumer rollout catches up, the model tier underneath is already callable — and a router with failover is the closest thing to an undo button an API has.

© 2026 OrcaRouter

For Providers

Run an inference platform? Get your models on OrcaRouter.

providers@orcarouter.ai

Join our community

Discordsupport@orcarouter.aiXGitHubYouTube