Gemini 4 Carbon arrives with a swarm of agents
Google opens its most powerful model to the public and previews in Antigravity a mode that turns one prompt into a team of coordinated agents.
One prompt and, on the other side, several coordinated agents working in your place. Google is previewing that feature, called "swarm," in Antigravity alongside Gemini 4 Carbon, its most powerful model, The Verge reported.
Two weeks ago, Carbon was reserved for a select group of "cybersecurity defenders." Now it is opening to paying developers and businesses through the Gemini API in Google AI Studio and Gemini for business. It is also coming to Google AI Ultra subscribers in the Gemini app, in AI Mode in Search and in Antigravity. For free or cheaper plans, there is no date or confirmation.
A swarm for long jobs
The feature targets projects that stretch out over time, the kind where a single agent falls short. You describe the goal and several instances split up the pieces. For someone who studies, starts a business or serves in a local church, the deeper change is in the shape of the work: it is no longer enough to write the instruction better. You have to know how to break the problem apart and review what comes back.
«Delegating execution is nothing new; what is new is the size of the team that responds.»
That is where many people stumble. The swarm delivers volume, not direction. If the instruction is vague, you multiply the vagueness by five.
Carbon, Argon and what it costs
Google maintains that Argon, launched last month, will remain the sensible choice for most software engineering and knowledge work. Carbon is for what demands more: complex multistep programming and open-ended scientific research.
The entry price is $2 per million input tokens and $10 per million output tokens. That is about one-fifth of OpenAI's GPT-6 Astra and one-fourth of Anthropic's Fable 5.2, which came out the day before. The introductory rate doubles at the end of November. The Verge cautions that a cheaper token does not always mean a smaller bill, because each model uses a different number of tokens for the same task.
Google published benchmarks that place Carbon among the most powerful models on the market, below Fable 5.2. The Verge reminds readers that companies tend to pick the results that favor them. With this move, Google returns to the forefront of AI research, after months in which the race seemed to belong to Anthropic and OpenAI.
What you hand off and what you keep
Delegating execution is nothing new; what is new is the size of the team that responds. You can ask a swarm to organize a spreadsheet or put together a draft of a report while you study. You can also have it try different paths on the same code. What you do not hand off is judgment: what it is for, whom it blesses or harms, what gets published and what stays unsaid.
Scripture does not separate the useful from the holy. "Whatever you do, work heartily, as for the Lord and not for men" (Colossians 3:23, ESV). Automating the repetitive can give you back hours for what only a person can do: listen to someone, care for a brother or sister, decide honestly. For discernment, see Proverbs 2:3-5.
This week, before asking an agent for something, write in one line what task you are handing off and under what condition you would accept the result; if you cannot describe the condition, do not hand it off. When you do it with a real task, you will know what was gained and what was lost. What would you give a swarm, and what would you never give it?
