AI Studio Ultra Mode: Smart, Surprising Google Build Leak

AI Studio Ultra mode is an unreleased option that Google is building into the app builder inside Google AI Studio, according to TestingCatalog’s report on 9 October 2026. It sits in the mode dropdown of AI Studio Build, next to Plan mode, Build mode and a Security review option that was itself spotted only nine days earlier. Its one-line description reads “Build with advanced skills and tools”. Google has not announced it, documented it or said who will get it.

That leaves a short list of facts and a longer list of open questions. TestingCatalog’s early tests suggest the new mode spends more effort on each app it generates. Nobody outside Google yet knows which skills and tools it adds, whether it will be tied to the Google AI Ultra subscription, or whether it was built for Gemini 4 Argon, the frontier model Google announced on 30 September.

This article sets out what was found, what Build mode already does, how AI Studio is priced and tiered today, and what the new mode could change for teams that prototype with Gemini. It follows our coverage of the Antigravity agent rename, because the same agent harness now powers Build mode.

Table of contents

What TestingCatalog Found: AI Studio Ultra Mode in the Build Menu
How AI Studio Ultra Mode Behaved in Early Tests
What Build Mode Already Does Before AI Studio Ultra Mode Arrives
Reading “Advanced Skills and Tools” in AI Studio Ultra Mode
Will AI Studio Ultra Mode Need a Google AI Ultra Subscription?
AI Studio Ultra Mode and Gemini 4 Argon
Why AI Studio Ultra Mode Is Not OpenAI’s Ultrafast
What AI Studio Ultra Mode Means for Business Builders
A Checklist Before You Switch On AI Studio Ultra Mode
Questions About AI Studio Ultra Mode
References

What TestingCatalog Found: AI Studio Ultra Mode in the Build Menu

TestingCatalog, which tracks unreleased features in AI products, published its report at 12:47 UTC on 9 October under the byline of Alexey Shabanov. Three minutes later it posted the find on X: “A new Ultra mode has been spotted in development on Google AI Studio Build.” By the time we read it that evening, the post had more than 36,000 views.

The report is short and careful. It says the specific capabilities of AI Studio Ultra mode “remain unclear”, that any link to Argon “remains unconfirmed”, and that “it is also unclear whether the mode will require an Ultra subscription”. Those caveats matter more than the headline, because a dropdown label in a test build is not a launch.

The four options in the Build mode dropdown

Two screenshots in the report show the full menu. One is the AI Studio Build start page, captured late on 8 October. The other is taken inside a project on 9 October. Both list the same four entries in the same order.

Mode	Description shown in the menu	Status on 9 October 2026	First reported
Plan mode	“Create a plan before making changes”	In development	30 Sep 2026
Build mode	“Make changes right away”	Live, the default	Documented by Google
Ultra mode	“Build with advanced skills and tools”	In development	9 Oct 2026
Security review	“Apply secure coding practices”	In development	30 Sep 2026

The start page around the menu is the familiar one. It reads “Build your ideas with Gemini” above a prompt box that says “Describe an app and let Gemini do the rest”, with an “I’m feeling lucky” button beside it. Chips underneath offer “Build an Android app”, Google Drive and Google Calendar.

From three Build modes to four in nine days

On 30 September at 23:06 UTC, TestingCatalog posted a screenshot of the same dropdown with only three entries: Plan mode, Build mode and Security review. Its post asked whether the new options would be powered by Gemini 4 Argon. AI Studio Ultra mode was not in that menu. It appears between Build mode and Security review in the 8 and 9 October screenshots.

So Google added a fourth option to an unreleased menu within nine days, while the first two additions were still unreleased. That pace fits a team working towards a single launch of several modes rather than shipping each one alone.

How AI Studio Ultra Mode Behaved in Early Tests

TestingCatalog says its testers could run apps through the new flow. “In our early tests, generations produced through this flow were quite comprehensive despite using Gemini 3.8 Flash,” the report says. On that basis it concludes that AI Studio Ultra mode “appears to devote more thinking effort to app generation than the standard flow”.

The same paragraph sets the limit on that finding. The tests “don’t show how the mode allocates reasoning or which additional tools it uses”. In other words, the output looked more thorough, but the mechanism is a guess.

Gemini 3.8 Flash, not a frontier model

The model named in the test matters. Gemini 3.8 Flash is Google’s fast, cheaper model, released on 2 September at an introductory API price of $0.75 per million input tokens and $3.75 per million output tokens. TestingCatalog’s launch report said Google credited the model’s results to additional reasoning steps and iterative tool calls, and that developers could choose a lower effort level when compute efficiency mattered more.

That gives a plausible reading of AI Studio Ultra mode that needs no new model at all: run the same Flash model at a higher effort setting, let the agent make more tool calls, and accept longer and costlier generations in exchange for more complete apps. It is a reading, not a fact. Google has not said how the mode works.

What the test screenshot actually shows

One of the report’s screenshots shows the test itself. The prompt was simply “generate testingcatalog ai news website”. The panel names Gemini 3.8 Flash as the model. On the right, the preview shows a polished news site with filters by lab and topic, a “Leak Radar” counter and a lead story headed “ChatGPT Web Beta Uncovers Native MCP (Model Context Protocol) Client Integration”.

That story is invented. It carries a box that says the flag was “Confirmed in client JavaScript bundle chunk-connectors-mcp.js”, and the masthead calls the site “Decompiled & Verified”. Nothing in TestingCatalog’s own news feed matches it. The model filled the page with confident placeholder copy because the prompt asked for a news site and gave it no real news.

This is not a flaw unique to AI Studio Ultra mode. Every app builder writes sample content. But a mode designed to produce more complete apps will also produce more complete-looking fake content, and that raises the bar for review before anyone publishes the result.

What Build Mode Already Does Before AI Studio Ultra Mode Arrives

To judge what “advanced skills and tools” might add, it helps to know what Build mode already includes. Google’s Build mode documentation, last updated on 20 August 2026, describes a fairly complete platform. It does not mention Plan mode, Security review or AI Studio Ultra mode, which confirms all three are unreleased.

Full-stack web apps and native Android apps

Build mode creates a full-stack environment by default. The client side is a web front end, with React as the default, and the server side is a Node.js runtime for secure API calls, database connections and npm packages. The agent installs packages it decides the app needs, and you can name specific packages yourself.

Android is the second target. Build mode can generate a Kotlin and Jetpack Compose project, run it in a browser-based emulator, install it on a physical phone through ADB in the browser, and publish it to the Play Store for internal testing. Google’s own example chip on the start page reads “Build an Android app”.

The Antigravity agent harness behind Build

The most important line in the documentation is about the engine. “The Antigravity Agent is the main AI functionality within Google Antigravity and now the core components of the agent harness is powering the Build mode experience in Google AI Studio,” it says.

The page lists what that harness brings: it keeps the context of earlier prompts and file states, manages dependencies across multiple files, and “verifies code updates to reduce hallucinations”. That is the same agent Google has been reworking in its Antigravity coding product, where TestingCatalog found signs of Argon support on 7 October. If AI Studio Ultra mode gives that agent more room to work, it inherits those features rather than adding a new system.

Integrations, deployment and GitHub sync

Build mode already connects to a long list of services, which is why the phrase “advanced tools” is hard to pin down. The documented features include:

Firebase Firestore and Firebase Authentication, provisioned automatically, including “Sign in with Google”.
Google Workspace APIs such as Gmail, Sheets, Docs, Drive and Calendar, with OAuth set up by AI Studio.
A secrets store, with your Gemini API key injected into the server-side runtime and never into client code.
Multiplayer apps, where the server-side runtime manages shared state.
Two-way GitHub sync with AI-generated commit messages and a side-by-side merge conflict viewer.
Deployment to Cloud Run, or a ZIP download to run the app anywhere.

Annotation mode lets you highlight part of the app’s interface and describe the change you want. The App Gallery offers projects to remix. For most prototypes, Build mode is already enough.

Reading "Advanced Skills and Tools" in AI Studio Ultra Mode

The description gives three clues. “Build” says it is still the app builder. “Skills” and “tools” are words Google and its rivals use for specific things, and each points to a different possible design for AI Studio Ultra mode.

Skills: packaged know-how the agent can load

In agent products, a skill usually means a packaged set of instructions, and sometimes scripts, that an agent loads when a task needs it. Google has been moving in that direction elsewhere. On 30 September we reported that Google may retire Gems and migrate users to Skills in the Gemini app. If AI Studio Ultra mode gives the Build agent a library of skills, it might load set procedures for common jobs, such as adding payments, building an admin panel or writing tests, instead of improvising each time.

Tools: more things the agent can call

Tools are the actions an agent can take. Build mode can already install npm packages, set up Firebase and call Workspace APIs. “Advanced tools” might mean extra capabilities such as running the app and checking its behaviour, searching the web for current library documentation, or editing larger codebases. TestingCatalog’s report does not say which, and we will not guess at a list.

Effort: more thinking per generation

The third reading is the one TestingCatalog’s tests support. AI Studio Ultra mode may simply spend more reasoning and more tool calls on each request. Gemini 3.8 Flash already has effort levels, and Google has been explicit that higher effort uses more tokens. A mode that turns effort up would explain why Flash produced “quite comprehensive” apps in the test.

These readings are not exclusive. A mode could raise effort and load skills and call more tools at the same time. What matters for planning is that every one of them raises the compute behind each app, which brings the question of who pays.

Will AI Studio Ultra Mode Need a Google AI Ultra Subscription?

This is the open question TestingCatalog raised twice. In its X post it wrote: “There is no sign that Ultra mode will require an Ultra subscription, or which tools and skills it will use.” The name invites the link, because Google AI Ultra is the top tier of Google’s consumer subscriptions, and AI Studio already ties usage to those tiers.

How AI Studio is tiered today

Google’s Google AI plans page for AI Studio, last updated on 18 August 2026, sets out three levels. Subscriptions “unlock paid models and higher rate limits” in the Playground and features like the code assistant in Build mode, with daily quotas that reset rather than roll.

Plan	US price per month	AI Studio usage (Google’s wording)	Model access in AI Studio
Free	$0	“Modest quota”	Basic limits and access
Google AI Pro	$19.99	“Higher quota”	Premium models such as Gemini Pro, Nano Banana and Lyria
Google AI Ultra	$99.99 (5x) or $199.99 (20x)	“Highest quota”	“Highest limits for prototyping, development, and advanced frontier models”

Google’s consumer plans page adds a row that bears directly on AI Studio Ultra mode. Under Google AI Studio it lists “Vibe Coding agent: Daily usage across Flash and Pro models” and grades it Limited, Expanded, Higher and Highest across Plus, Pro, Ultra 5x and Ultra 20x. The Build agent’s daily allowance already scales with the plan, so a mode that burns more effort per app would use that allowance faster on every tier.

What each plan costs per unit of usage

The US plans page states each tier’s headline usage as a multiple. AI Plus gives 2x the Free limits, AI Pro gives 4x, and the two Ultra plans give 5x and 20x the limits of AI Pro, which is 20x and 80x Free. Dividing each price by its multiple of Free gives a rough price per unit of usage.

Monthly price per “Free-equivalent” unit of usage (US dollars, price divided by usage multiple)

Google AI Plus ($4.99 / 2) $2.50

Google AI Pro ($19.99 / 4) $5.00

Google AI Ultra 5x ($99.99 / 20) $5.00

Google AI Ultra 20x ($199.99 / 80) $2.50

Two caveats apply. The multiples are Google’s headline figures for Gemini app usage, while AI Studio’s own limits are described only in words. And the arithmetic ignores everything else in each bundle, such as storage and YouTube Premium. Still, the shape is clear: the top Ultra tier is the cheapest way to buy heavy usage, which is exactly what a high-effort build mode would consume.

The agent exception in Google’s AI Studio rules

One line on the AI plans page cuts the other way. “Access to agents (Deep Research and Antigravity Preview) within AI Studio is not included in Google AI plans and requires a paid API key,” it says. Subscriptions also apply “only within the Google AI Studio web interface”, and Google One AI credits are not accepted at all.

So Google already runs two payment routes inside AI Studio. Build mode’s daily allowance comes from your subscription, while some agent features need a billed API key. If AI Studio Ultra mode leans on the full Antigravity agent, Google could put it on either route, or offer both. That is a pricing decision, and it is not ours to predict.

AI Studio Ultra Mode and Gemini 4 Argon

TestingCatalog’s X post went further than its article: “these new modes are likely being developed to work with Gemini 4 Argon.” The article itself was more cautious, saying AI Studio Ultra mode “could arrive alongside” Argon’s wider release but that the link is unconfirmed.

Argon’s stated release order

Google’s Argon announcement on 30 September gave an order of access, though no dates. Argon is “rolling out to trusted cyber defenders through the Fairwind Program”, which reflects how much of its launch Google built around cybersecurity work. After that cohort’s feedback, Google plans to release it “to developers, enterprises, and consumers, starting with paid API customers and Google AI Ultra subscribers.”

That sentence is the strongest circumstantial link. The first two groups to get Argon are paid API customers and Ultra subscribers, which are the same two payment routes that already exist inside AI Studio. A Build mode called Ultra, tuned for long agentic runs, would be a natural place to put a model whose output limit Google raised from 64,000 to one million tokens. We covered the model’s benchmarks and pricing when Google first tested Gemini 4 Argon.

What a long build costs on each model

If AI Studio Ultra mode does move to Argon, the model price changes the economics. Argon’s introductory API price is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period. Gemini 3.8 Flash costs $0.75 and $3.75.

Output cost of an illustrative 300,000-token app generation (list API prices)

Gemini 3.8 Flash (300,000 x $3.75 per million) $1.13

Gemini 4 Argon, introductory (300,000 x $10 per million) $3.00

Gemini 4 Argon, standard (300,000 x $20 per million) $6.00

The 300,000-token figure is an illustration, not a measurement; nobody has published the token use of an AI Studio Ultra mode run. The ratios are the point. At standard prices, Argon output costs about 5.3 times Flash output, so the same long generation that costs about a dollar on Flash costs about six on Argon. Inside a subscription, that cost shows up as a faster-draining daily allowance rather than a bill.

Why AI Studio Ultra Mode Is Not OpenAI's Ultrafast

TestingCatalog ended its report with a warning: “there is little reason so far to treat it as equivalent to OpenAI’s Ultrafast mode.” The similar names make the comparison tempting, but the two products pull in opposite directions.

Question	Google AI Studio Ultra mode	OpenAI Ultrafast
What it changes	Depth: “advanced skills and tools”, apparently more effort per app	Speed: faster token output from the same model
Where it lives	The app builder in AI Studio	ChatGPT, Codex and the API
Price	Not announced	API at 6x Standard prices; included usage burns at 8x
Who gets it	Not announced	Pro $500, plus eligible Enterprise and Edu
Status	Unreleased, seen in test builds	Live since 8 Oct 2026 for GPT-6.1 Sol

Ultrafast is a speed premium on a model you already use, and we set out its rates when OpenAI rolled out GPT-6.1 Sol Ultrafast. AI Studio Ultra mode, on the evidence so far, is the opposite trade: wait longer and spend more compute for a more complete app. Both are ways of selling more compute to people who will pay for it, which is the only real thing they share.

What AI Studio Ultra Mode Means for Business Builders

For a business, AI Studio is mostly a prototyping tool. Teams use it to turn an idea into a working demo in an afternoon, then decide whether it deserves real engineering. A deeper build mode makes those demos more convincing, which is both the benefit and the risk.

Faster prototypes, with the same review duty

A more complete first draft saves time if it is right. Our guide to using AI coding assistants safely in large software projects makes the point that generated code still needs the same review as hand-written code. That applies twice as strongly when the output looks finished. TestingCatalog’s test app looked ready to publish, and its lead story was fiction.

Keys, costs and Cloud Run deployments

One documented behaviour deserves attention before anyone deploys a deeper build. When you deploy to Cloud Run from AI Studio, “the deployed app will use your API key for all users’ Gemini API calls”. Google keeps the key on the server, which is good security practice, but every visitor’s AI call is billed to you.

If AI Studio Ultra mode produces apps that make more Gemini calls, or calls to a costlier model, a public demo can run up a real bill. Set API budgets and alerts in Google Cloud before sharing any app built this way, and keep demos behind sign-in where possible.

Security review should be a habit, not a mode

It is telling that Google is building a Security review option at the same time as AI Studio Ultra mode. A mode that applies “secure coding practices” is welcome, but it suggests the default flow does not always do so. Treat it as one check among several, not as a guarantee. Our earlier look at the rise and risk of vibe coding covers the wider problem of apps that work in a demo and fail under real users.

Who owns what the agent builds

Google’s documentation is clear that the developer carries the legal duties. Its list of developer responsibilities includes compliance with “all applicable laws and regulations”, content monitoring under the Google Cloud terms, and “safeguards and moderation tools to prevent your application from being misused”. A smarter build mode does not move any of that onto Google.

A Checklist Before You Switch On AI Studio Ultra Mode

When AI Studio Ultra mode reaches your account, a short routine will keep the gains and limit the surprises. None of these steps depends on details Google has not yet published.

Before the first run

Check which model the run uses. The Build panel names it, as TestingCatalog’s screenshot shows.
Check your plan and daily allowance, and whether the mode needs a paid API key.
Write the prompt with real data or clearly marked placeholders, so invented content is easy to spot.

Before you share or deploy

Read every piece of generated copy, not just the code. Delete anything that states a fact you did not supply.
Run Security review, if it is available, and then review authentication, secrets and data storage yourself.
Set a Google Cloud budget alert before a Cloud Run deployment, because users’ calls bill to your key.
Sync to GitHub so changes are versioned and reviewable outside AI Studio.

Before it becomes a product

Move from subscription quota to a production project with Cloud Billing, which Google itself recommends for scale.
Assign an owner for the app, its data and its costs.
Plan for a model change, since the mode may switch models as Argon rolls out.

Questions About AI Studio Ultra Mode

Is AI Studio Ultra mode available now?

No. As of 9 October 2026 it has only been seen in test builds by TestingCatalog. Google’s Build mode documentation, updated in August, does not mention it, and Google has made no announcement.

What does AI Studio Ultra mode do?

Its menu description is “Build with advanced skills and tools”. TestingCatalog’s tests suggest it puts more thinking effort into each app, producing more complete results even on Gemini 3.8 Flash. The specific skills and tools are not known.

Will it require a Google AI Ultra subscription?

Unknown. TestingCatalog found “no sign” that it will. AI Studio already gives Ultra subscribers the highest Build quotas, and Google says Argon will reach Ultra subscribers and paid API customers first, so a link is plausible but unconfirmed.

Is it powered by Gemini 4 Argon?

Not in the tests seen so far, which used Gemini 3.8 Flash. TestingCatalog believes the new modes are “likely being developed to work with” Argon, but Google has not confirmed that.

How is it different from Plan mode and Security review?

Plan mode creates a plan before making changes, and Security review applies secure coding practices. AI Studio Ultra mode is described as building with advanced skills and tools, which suggests a more capable version of Build mode rather than a planning or checking step.