On November 10, 2026, Google expanded availability of its AI model Gemini 4 Carbon to paid developer and enterprise APIs and to Google AI Ultra subscribers. The model had been in limited release since October, and Google also made clear that the introductory API pricing will end on November 30.
Google also announced a preview of "/swarm," a new feature that lets multiple AI agents split up development work and run it in parallel. The API bills by token usage, while Google AI Ultra draws on the usage allowance included in the subscription. When combining multiple agents, which tasks you assign to Carbon, and how many tokens or how much of your allowance the whole job consumes, will be key factors in judging whether adoption pays off.
Gemini 4 Carbon expands to the paid API and Google AI Ultra
The paid API for Gemini 4 Carbon is offered through Google AI Studio and Gemini for business. For Google AI Ultra subscribers, it will roll out gradually through Antigravity, the Gemini app, and AI Mode in Google Search. In its official announcement on November 10, Google also lists Android Studio and CodeMender as places where it can be used. However, Google does not state that it is immediately available in every region, including Japan, or for every account, so actual availability needs to be checked for each environment.
In the October 28 announcement, the model was offered only to a limited group of cyber defenders through the "Fairwind Program." Developers and enterprises can now use it through the paid API, but the advanced capabilities and evaluation results of "Carbon for cyber," designed for cyber defense, do not necessarily apply to the general-purpose API. Expanding who can use the model is distinct from opening up features built for a specific field.
In Antigravity, the environments where the model itself and the new "/swarm" feature are available also differ. Gemini 4 Carbon is available to Google AI Ultra subscribers in Antigravity's App, CLI, and IDE extension, but the /swarm preview is limited to the App and CLI. Google has not said that /swarm can be used in the IDE extension. The introductory pricing through November 30 is a condition that applies to the API for the generally available Carbon, and does not apply as-is to Google AI Ultra usage allowances or to the terms of the Fairwind Program.
There is also a separate support program for scientific research. About four weeks before this announcement, Google began accepting applications for a program offering Gemini 4 Argon free for six months to 10,000 scientists, and says it has already admitted several thousand. Current participants and researchers who join later will additionally be given access to Gemini 4 Carbon. However, the program has eligibility requirements and a selection process, and participants are admitted in phases each week, so it is not open unconditionally to every scientist.
/swarm lets multiple AI agents divide up development work
The newly announced "/swarm" is a feature that lets multiple AI agents take on different roles and carry out software development in parallel. A coordinating agent launches specialized worker agents, each running in its own independent Git worktree. This allows changes to the same repository to proceed in parallel in separate working locations. It also includes a verification agent that runs tests, reviews code diffs, and checks for implementation problems.
According to the official Antigravity team blog, each agent can exchange information directly through an append-only shared bulletin board. Rather than routing all communication through the coordinator, the design lets agents share progress and needed information with one another. By combining isolated working environments with inter-agent communication, the setup lets multiple tasks move forward at once while the final deliverable is assembled.
One notable feature is that you can specify which AI model to use for each task. You do not have to assign Gemini 4 Carbon to every agent. You can use Carbon for complex design, overall coordination, cybersecurity audits, and verification, while using lighter, faster models for ordinary code editing and implementation. Separating work that requires advanced reasoning from relatively simple processing makes it possible to match the model to the nature of the task.
Google itself says Gemini 4 Argon is well suited to most software development and knowledge work. Carbon, by contrast, is intended for complex multi-step coding and debugging, open-ended scientific research, and similar tasks. It is designed to use more tokens than Argon for exploration, reasoning, and verification, and Google is not claiming that switching all work to Carbon is optimal. Rather, the key idea behind mixing models is to concentrate Carbon on the steps that require difficult judgment.
Google says it tested /swarm internally for five weeks. This is Google's own validation, however, not a result independently confirmed by a third party. The company also says agents consume no tokens while idle, but that does not eliminate consumption while they work. When running multiple agents in parallel, you need to track usage across the whole system rather than for each individual agent.
API prices double from December, and total cost of parallel runs is the issue
The general API pricing for Gemini 4 Carbon changes after November 30, 2026. According to the pricing terms Google presented in its November 10 announcement, the introductory price is $2 per million input tokens and $10 per million output tokens. From December 1, these become $4 and $20, respectively, doubling the unit price for both input and output.
| Billing item | Through Nov. 30, 2026 | From Dec. 1, 2026 | Change |
|---|---|---|---|
| Per 1M input tokens | $2 | $4 | 2x |
| Per 1M output tokens | $10 | $20 | 2x |
For the same number of tokens, the price from December 1 is twice the introductory price. The math is 4 ÷ 2 = 2 for input and 20 ÷ 10 = 2 for output, assuming the same model and unchanged consumption. In real work, however, token consumption varies with how the model reasons and how many retries occur, so the total bill for a job will not always be exactly double.
When running multiple agents at once, you also need to watch usage-allowance consumption, not just API fees. In its official blog, the Antigravity team offers this caution:
Parallel multi-agent systems consume usage allowances far faster than ordinary chat.
Even if you eliminate token consumption during idle time or hand simple processing to lighter models, total consumption across the system may still rise as more agents do work. Google has not published how to convert Google AI Ultra usage allowances into API pricing, nor details of the allowance available for /swarm. Nor has it given specific figures on what percentage of cost can be saved by assigning models by role, or which combinations are optimal.
For that reason, what to compare when adopting the model is not just the price of a single response. You need to evaluate the total tokens consumed until the development work is finished, the time required, the quality of the code and other deliverables produced, and the time humans need for review and correction. If Carbon reasons with more tokens and that reduces later fixes and reruns, a higher unit price could still come out ahead across the whole job. Conversely, if output quality barely differs, using a lighter model may make more sense. Which is better can only be determined by running the same task and comparing.
Second place in quality on the MKA benchmark, at roughly one-sixth the cost of the top model
To evaluate how well AI models handle complex knowledge work, Google is developing the "Multimodal Knowledge Agent (MKA)" benchmark together with Artificial Analysis. According to the official explanation of the benchmark, it measures the ability to read drawings, charts, and document page layouts and complete a deliverable through multiple steps. Experts across 10 fields and 83 occupations created a total of 200 tasks, targeting work that takes skilled humans three to eight hours. Its distinguishing feature is that it evaluates the ability to complete complex, practice-like work, rather than the ability to answer short questions correctly.
Scoring uses about 70 yes/no criteria per task. Points are weighted 75% for factual accuracy and 25% for visual design, and all models run in the same "Stirrup" harness, with the same tools, execution environment, and budget of response turns. Google says agreement between automated grading using non-Google models and expert evaluation was 96%. That figure indicates the consistency of scoring results, however, and does not mean AI can replace 96% of human work.
According to Google, Gemini 4 Carbon ranked second in deliverable quality on MKA, with an average cost per task of roughly one-sixth that of the top-ranked model. This can be read as showing a degree of competitiveness in the balance between quality and cost under the same evaluation conditions. However, the cost difference is based on the execution conditions set in the benchmark, and the same gap will not necessarily appear in general development work or enterprise operations.
The 200 tasks used in evaluation are not public, with a separate set of 10 tasks released. Results should also be interpreted with the fact in mind that MKA itself is an evaluation developed jointly by Google and Artificial Analysis. While it is a useful reference in that it measures multi-step tasks close to real work, it does not directly guarantee performance in any particular company's development environment or operations.
The Gemini 4 Carbon model card lists limitations that include generating incorrect information, response delays and timeouts, and higher token consumption at high reasoning settings. Its knowledge cutoff is March 2026, and in some fields it may only have older information. For work involving the latest software specifications, research findings, or changes in the external environment, content generated by the model needs to be verified separately.
When considering Gemini 4 Carbon, it is important to compare the cost and quality of completing actual work, not just benchmark rankings or the API price per request. With /swarm in particular, you will want to separate the steps assigned to Carbon from those left to lighter models, and measure deliverable quality, verification burden, completion time, and total token consumption. Calculating costs under both the introductory pricing through November 30 and the standard pricing from December 1 will help identify which tasks are still worth running on Carbon after the price change. The key to the decision is whether you can pinpoint the steps where advanced reasoning actually improves deliverable quality or development efficiency.