Google released Gemini 4 Carbon more broadly on November 10, 2026, following a two-week trial with a selected group of cybersecurity defenders. The Artificial Intelligence (AI) model is aimed at paying businesses, developers, and Google AI Ultra subscribers. For you, the key points are its actual availability, the higher prices coming in December, and a preview of the agent swarm in Google Antigravity.
Availability with a clear paywall
Gemini 4 Carbon began rolling out on November 10 through the Gemini app, AI Mode in Google Search, and Google Antigravity for all Google AI Ultra subscribers. Paying businesses and developers also receive access through Google AI Studio and Gemini for Business, according to the Google DeepMind announcement. An application programming interface (API), which lets other software connect to the model, is available there as well.
This is not a general free release. Google has not provided a date for lower-priced subscriptions or free accounts, nor has it confirmed that Carbon will become available through those tiers at all. The supplied sources also do not state the subscription price of Google AI Ultra.
Carbon does not replace Gemini 4 Argon. Google continues to recommend Argon for most knowledge work and software engineering, while Carbon is intended for unusually difficult, multistep assignments. Its examples include complex debugging, longer coding projects, and open-ended scientific research in which the model needs to examine several possible solutions.
The sources provide no Switzerland-specific details about regional access, supported languages, or data-processing terms. Google’s statement that Carbon is rolling out to all Ultra subscribers sounds broad, but it is not an explicit confirmation for Swiss accounts. The five reports also lack enough information about storage location, retention, or the use of prompts to assess professional work with confidential documents.
Prices that tell only part of the story
At launch, Carbon costs $2 per million input tokens and $10 per million output tokens. A token is a small unit of text used to process prompts and generate answers. According to Simon Willison’s launch overview, those rates will increase on December 1 to $4 and $20, respectively.
Google is initially charging the same per-token rates for Carbon and Argon. Carbon, however, uses more tokens, according to the company, so it can explore alternatives more thoroughly and run longer verification cycles. A demanding Carbon task may therefore cost more overall even when the listed rate for each million tokens is identical.
The same issue affects comparisons with competing models. The Verge describes the introductory rates as considerably lower than those of certain flagship models from OpenAI and Anthropic, but it also notes that different models may consume different numbers of tokens for the same assignment. A low rate on a price sheet does not by itself guarantee a lower final bill.
An Artificial Analysis test covered by The Decoder calculated a cost of $7.76 per solved task for Carbon using its regular price. GPT-6 Astra Max came to $6.25, Opus 5.5 Max to $10.04, and Fable 5.2 at its highest setting to $26.50. A lower Fable configuration cost about $9 per task while still finishing two points ahead of Carbon in the overall index.
These results do not necessarily contradict every performance claim, but they qualify Google’s selection of favorable results. Google reports first-place finishes on CursorBench 4.0, WANDR, and AutomationBench at more than twice the cost efficiency of the next-best external model. Benchmarks are standardized performance tests that cover different workloads, and providers can select the results that present their products most favorably; Carbon has therefore not been independently established as the best model for every purpose.
Agent swarms and Android features to try
The most prominent new feature is a preview of /swarm in Google Antigravity. An AI agent is a system that breaks a goal into smaller steps and carries them out with a degree of autonomy. The swarm is designed to turn one prompt into several coordinated agents whose activity you can monitor in real time.
Google describes an arrangement in which Carbon handles demanding roles such as planning, coordination, and review. Faster, lighter models can complete individual tasks in parallel, isolated workspaces, while additional agents critically test their changes. This setup is more relevant to a large project than to summarizing a short document or rewriting an email.
One example shows how far that division of labor can go. In the demonstration, /swarm created an open-world 3D game set on Google’s Mountain View campus. Subagents produced 3D assets in Blender, assembled a browser game with Three.js, and added sound effects and music. It is an impressive demonstration, but not independent evidence that arbitrary projects will run just as smoothly.
Simon Willison also tested Carbon with a request for a 3D animation in which a large machine walks through Manhattan, Queens, and Brooklyn using OpenStreetMap data. He had to follow up with an explicit request before receiving a self-contained HTML file without external dependencies. That detail is useful in practice: even a capable model will not necessarily deliver the exact final format on its first attempt.
On the same day, Google announced more Android features under the Gemini Intelligence name. The announcement does not establish that Carbon directly powers every one of these features, but it is part of the wider Gemini product rollout. According to Google’s Android announcement, Pixel 10 models are among the devices receiving new capabilities, while the Motorola Signature 27 is scheduled for U.S. preorders starting November 19.
Rambler turns spoken thoughts into organized text in real time. Task automation can handle activities such as ordering food, hailing rides, and making reservations across apps, and it is expanding to the United Kingdom, India, and Japan rather than explicitly to Switzerland. Create My Widget will also let you describe a travel dashboard, monthly spending tracker, or countdown in ordinary language; it is rolling out over several weeks and can incorporate information from Gmail, Calendar, and Search.
Pros and Cons of Gemini 4 Carbon
Pros:
- Demanding assignments – Carbon is designed specifically for long, multistep work that benefits from deeper verification.
- Coordinated agents – The swarm can distribute planning, production, and review across several specialized systems.
- Low introductory rates – Initial token prices are below the regular December rates and several cited competing offers.
- Multiple access routes – Ultra subscribers can use Carbon in Gemini, Search, and Antigravity, while businesses gain additional access through Google’s work platforms.
Cons:
- No free release – Google has provided neither a date nor a firm commitment for free or lower-priced accounts.
- Higher consumption – Carbon uses more tokens than Argon, which can increase the total cost despite matching rates.
- Unproven swarm preview – The striking agent examples come from Google or individual early tests and do not yet demonstrate consistent reliability.
- Open Swiss questions – The reports do not specify regional availability, language support, or data-protection conditions for Switzerland.
A practical way to get started
If you are a beginner, Carbon makes the most sense when your assignment genuinely requires several stages and checks. A focused research task, an extended document comparison, or a structured project plan will reveal more about its value than a short everyday question. Watch whether the answer remains traceable and whether Carbon generates more text than the assignment requires.
The Android features may offer a more approachable everyday starting point if they reach your device and region. A travel widget built from Calendar and Search information, or Rambler organizing your spoken notes, is easier to evaluate than a complex agent swarm. When a feature accesses Gmail, Calendar, or other personal information, you should inspect its displayed permissions and results individually.
If you are an advanced user, you can compare Argon and Carbon on the same substantial assignment. Relevant measures include not only output quality and listed token rates, but also consumption, required corrections, and the number of failed attempts. In Antigravity, a clear division of roles is useful: Carbon can handle planning and review, while lighter models perform bounded tasks in parallel.
The swarm should not be mistaken for an independently accountable team. If several agents build on the same false assumption, they may complete the work faster without making it more accurate. Results involving research, data analysis, or business decisions therefore still require expert review outside the system.
Gemini 4 Carbon expands Google’s offering mainly at the high end, with more computation, longer verification, and its first tangible preview of coordinated agents. The introductory rates are attractive, but the December increase and Carbon’s greater token consumption make the real cost less straightforward. The main unresolved issues are the swarm’s reliability and specific availability and privacy terms for Switzerland.
Sources
- Google releases Gemini 4 Carbon with agent ‘swarm’ feature – The Verge, 2026-11-10
- llm-gemini 0.36 – Simon Willison’s Weblog, 2026-11-10
- Google startet Gemini 4 Carbon mit neuer Agenten-Schwarm-Funktion – The Decoder, 2026-11-10
- Gemini 4 Carbon is now available – Google DeepMind, 2026-11-10
- Gemini Intelligence: New features, on more devices – Google, 2026-11-10
Image: MART PRODUCTION via Pexels