Connect with us

NEWS

Google Puts Gemini 4 Carbon Behind the Ultra Gate

Google is sending Gemini 4 Carbon to AI Ultra in the Gemini app, keeping Argon as the daily model as API rates double December 1.

Published

on

Google opened Gemini 4 Carbon to AI Ultra users in the Gemini app on November 10, while leaving Argon as the default for most work. Paying developers can call the same model through the Gemini API at $2 per million input tokens and $10 per million output tokens through November 30, a rate that doubles on December 1.

Koray Kavukcuoglu, SVP at Google DeepMind and Google’s chief AI architect, wrote that Argon remains the ideal choice for the majority of software engineering and knowledge work, and that Carbon is built for tasks that test the limits of a model’s intelligence. The Gemini app picker now labels 4 Argon as “Code & analysis” and 4 Carbon as “Most complex work,” and Gemini 3.1 Pro is gone from that list.

Gemini 4 Carbon Lands Behind an Ultra Gate

The November 10 post is the public half of a rollout that started on October 28, when Google announced Carbon as its strongest model and sent it first to a subset of cyber defenders. Kavukcuoglu’s availability note says Carbon is now available to developers, enterprises, scientists, and consumers, and that Carbon uses more tokens than Argon to explore, reason, and run extra check loops.

The DeepMind model card states Carbon is based on Gemini 4 Argon, with a combined input-output context window of up to 1 million tokens. That is the same long window as Argon. The split is about how hard the job is, and who is allowed to pick the heavier model in a consumer app.

Google’s own phrasing on X is blunt: if Argon is great for complex tasks, Carbon is the model you bring in for your hardest problems. That line is the product. A new SKU for peak work sits next to a cheaper daily brain, and the app copy tells people not to reach for Carbon first.

Who Can Use Gemini 4 Carbon

Google AI Ultra users can pick Carbon in the Gemini app, AI Mode in Search, Google Antigravity, and Google AI Studio, while any paying API customer can call gemini-4-carbon in AI Studio and Android Studio. Consumer Pro plans keep Argon. Carbon has no free API tier, and thinking tokens are billed as output.

WHERE CARBON SHOWS UP

Channel Who gets it Carbon access
Gemini app and AI Mode Google AI Ultra Yes, from November 10
Google Antigravity Google AI Ultra Yes, including /swarm preview
Gemini API and AI Studio Paid API users Yes, model id gemini-4-carbon
Gemini for business Paying enterprises Yes, via API or model menu
Google AI Pro in the app Pro users Argon only
Scientists and researchers Application, rolling Both models, 6 months, highest Antigravity tier
Cyber defenders Vetted partners Carbon for cyber, rolling

Google’s U.S. plans page lists Google AI Ultra plan benefits from $99.99 a month, with higher Gemini limits than Pro and a Carbon line on the Ultra card. Pro still gets Argon, which Google still recommends for most coding. That same Gemini 4 line already finished a long Minecraft run in multi-agent tests, which is the kind of grind Google now wants people to keep on Argon.

The app lock is the choice users will feel first. Rival labs put their top model on more consumer plans. Google did not. Anyone who will not pay Ultra can still buy Carbon by the token, which is a different product with a clock on the sticker.

The $2 Token Price Holds Through November 30

Carbon launched at Argon’s launch price. Google AI Studio’s pricing page lists the introductory $2 and $10 rates through November 30, 2026, then $4 per million input tokens and $20 per million output tokens starting December 1. Cached input is 95% off, or $0.10 per million through November 30 and $0.20 after that.

GEMINI 4 TOKEN PRICES

Meter, paid tier Through November 30 From December 1
Input, per 1M tokens $2 $4
Output, including thinking, per 1M tokens $10 $20
Cached input, per 1M tokens $0.10 $0.20

Every token the model processes is billed, including thinking tokens and internal tool instructions when grounding is on. Carbon is built to spend those thinking tokens. A low list price can still produce a high bill on a long agent run, and a pipeline priced on the November meter will pay double in December.

Google also claims Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, at more than 2x lower cost per task than the next best model from another company. Cursor’s own note puts Carbon at 60.4% on CursorBench at high effort, the top mark on that board. Those are coding and research agent scores, not a reason to point every chat at Carbon.

Second Place on Google’s Own Office Test

Existing tests, Google says, miss how agents read dense files with charts, images, blueprints, schematics, and tables. With Artificial Analysis it launched MKA, the Multimodal Knowledge Agent benchmark: 200 private tasks across 10 verticals and 83 jobs, each built as a three-to-eight-hour professional project with a rubric of about 70 yes/no checks, weighted 75% on facts and 25% on visual design.

Google’s own MKA write-up ranks Carbon #2 in quality among the models it tested, and says Carbon finishes tasks at 6x lower average cost than the top scoring model. The independent Multimodal Knowledge Agent leaderboard tells the quality race in Climb WSR, the harder slice of each rubric.

MKA CLIMB WSR

  • Claude Fable 5.2: 57%, first on the Artificial Analysis board, and first in five of the 10 verticals.
  • Gemini 4 Carbon: 54% at high thinking, second overall, first in healthcare.
  • Claude Opus 5.5: 53% at max, third overall, first in education and construction.
  • Gemini 4 Argon: leads the science vertical on the same board.

No model leads all 10 verticals. Word and plain-text outputs are the hardest file types; no model clears 52% on either. Google also released 10 public sample tasks on Hugging Face, and says domain experts agreed with the non-Google autorater 96% of the time. When Google first announced Carbon on October 28, it said the model ranked first on Artificial Analysis Intelligence Index v5.0 at 59. That index is a different test from MKA.

One Prompt Becomes a Swarm in Antigravity

For bigger builds, Google is previewing /swarm in Antigravity, which turns one prompt into a coordinated team of agents you can watch in real time. Developers can put Carbon on design, coordination, and audit roles, then pair it with faster, lighter models in parallel workspaces and with adversarial checkers that test every change.

JOBS GOOGLE ASSIGNS TO CARBON

  • Hard engineering: Multi-package migrations, cross-subsystem root-cause work, and complex redesigns that need several agents.
  • Web apps: Build-and-verify loops that recover from errors and check both the code and the rendered UI.
  • Office artifacts: Spreadsheets, decks, and reports that follow strict schemas across dense PDFs, charts, and tables.
  • 3D and games: Stronger visual design and spatial reasoning on creative coding jobs.
  • Science and cyber: Open-ended research, plus find, exploit, and patch work for vetted defenders.

One Antigravity demo used seven public city-council planning PDFs, 206 extracted engineering drawing views, and a single prompt. A multi-agent Carbon team built an interactive 906K-triangle reconstruction of Google’s Platform 37 office across 10 overnight loops. Another pass rebuilt a Google merchandise storefront in two prompts, with new product shots from Nano Banana 2.1.

When we struggled to get our app looking consistent with an existing UI framework, we gave Carbon a list of requirements before closing our day. Overnight, Carbon had built a new Rust UI library from scratch and rewritten our OpenCode desktop app with working demos. It’s an incredibly brainy model and now a preferred planning model for our hardest tasks.

Kit Langton, Software Engineer, OpenCode, Google Carbon launch post

Baki Celtik, an AI engineer at HubX, said Carbon hit a 94% resolution rate across 16 real bug fixes in a 14,000-test codebase, with zero regressions. Shanks Salwan, lead applied AI engineer at Hebbia, called it the strongest model his team had evaluated for deep financial research, citing cleaner Excel models and sharper PowerPoint decks.

Fairwind and a Six-Month Lab Pass

Carbon’s first users were not Ultra subscribers. They were cyber defenders. Google is still routing the strongest cyber skills through a vetted track, while labs can apply for a time-limited pass on Antigravity’s highest usage tier.

THE CARBON ROLLOUT CLOCK

  1. October 28, 2026: Google announces Gemini 4 Carbon, cites an Intelligence Index score of 59, and opens the model to a subset of cyber defenders.
  2. November 10, 2026: Carbon reaches the paid Gemini API, Gemini for business, and Google AI Ultra surfaces in the Gemini app, AI Mode, Antigravity, and AI Studio.
  3. November 30, 2026: Last day of the $2 / $10 launch meter, with cached input at $0.10 per million tokens.
  4. December 1, 2026: Paid-tier input and output prices double to $4 and $20 per million tokens.

On the October 28 benches, run without guardrails, Google put Carbon at 88.2% on SEC-bench Pro for finding flaws, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. In Project Fortify, Google’s internal scan-triage-fix stack, Carbon hit 100% recall on scanning and 95% precision on triage. Google also says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, and that it resisted every evaluated attack on Gray Swan’s indirect prompt-injection test, after Argon had a 0.7% attack success rate.

Wiz, using Carbon in its Scan for Good program, said the model uncovered 0-day flaws at two major SaaS providers in under two hours. Axiom Bio, a drug toxicity research firm, said Carbon found toxicity mechanisms the team had not seen, and is moving production scientific analyses onto Carbon and Argon. In October, Google opened applications for 10,000 scientists to get Antigravity’s highest tier free for 6 months; on November 10 it expanded that offer so current and future participants can use Carbon as well, subject to eligibility, capacity, and selection. Cyber teams still apply through the Fairwind Program for cyber defenders.

Google Still Tells Most Users to Pick Argon

The unusual part of this launch is not another frontier score. It is Google shipping a stronger model and telling most customers to keep using the previous one. Carbon is the specialist. Argon is the default. Ultra is the consumer key for the specialist in the Gemini app, and the API meter is the other door, with a double on December 1.

On MKA, Carbon is the cheaper high-end office agent in Google’s telling, not the quality leader. On coding boards it can sit first and still be the wrong default for a short prompt. Teams that want the new model have a short window at $2 and $10 to measure cost per finished task against Argon, then a higher bill if the extra thinking is worth it.

The picker in the Gemini app now makes that choice visible. Code and analysis on one side, most complex work on the other, and 3.1 Pro no longer in the middle.

Harry is the editor of BUDGY APP, an independent title he owns and runs after ten years in journalism that began on a reporter's desk and ended up at the editor's. Numbers get particular attention here. A percentage in a business story is recomputed from the underlying figures before it goes live, a benchmark in a technology or gaming review is quoted with the conditions it was measured under, and a transfer fee or a lap time in the sports and auto pages is traced back to the club, the league or the timing sheet that published it. The same rule covers news, science, entertainment, lifestyle and travel: if a figure cannot be tied to a filing, a dataset, a transcript or a test Harry ran himself, it does not appear. Readers around the world see prices in the original currency with a conversion alongside. Errors are corrected in the open under a published corrections policy, with the change noted on the article. Questions about any figure reach him at support@budgyapp.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending