Google broadens Gemini 4 Carbon rollout as it tests a multi-agent ‘swarm’

Google is widening access to Gemini 4 Carbon, its newest frontier model, while also previewing an Antigravity “swarm” feature that breaks a single prompt into multiple coordinated agents. The move matters because it pairs a cheaper, more widely available model with a more ambitious workflow tool as Google tries to press its case against Anthropic and OpenAI.

What Google launched

According to The Verge, Carbon is now rolling out beyond the small cybersecurity-focused group that got the first release. The model is reaching users through the Gemini API, AI Studio, Gemini for business, the Gemini app for AI Ultra subscribers, Search AI Mode, and Antigravity.

That broader distribution marks a shift from the initial launch, which The Verge said was limited to selected cybersecurity defenders and professionals. The New Indian Express described the same model as a system built for coding, cybersecurity, and especially complex work that can stretch over days.

How does Carbon fit into Google’s model strategy?

Carbon is Google’s latest frontier model in the Gemini line, following Argon, which The Verge said arrived last month. The New Indian Express said Carbon builds on Argon and was announced two weeks later, underscoring how quickly Google is iterating at the high end of its AI stack.

Google’s pitch is familiar in the frontier-model race: better performance on hard tasks, stronger coding ability, and lower cost. But Carbon is not being positioned only as a model for chat or quick answers. Google says it is aimed at “tasks that test the limits of a model’s intelligence,” and The Verge reports that the company sees it as suitable for long-horizon, intelligence-heavy work such as multi-step coding and open-ended scientific research.

Why it matters

Carbon’s rollout matters for three reasons. First, it gives more developers and enterprise users access to Google’s newest flagship model after a restricted start. Second, it signals that Google is trying to compete not just on quality, but on price. Third, it adds a new workflow layer in Antigravity that could change how users think about AI agents.

The competitive context is plain. The Verge said Google considers Carbon its strongest model yet and believes it can stand alongside top systems from Anthropic and OpenAI. The same report said the launch marks Google’s return to the front line of AI research after months in which Anthropic and OpenAI have dominated the conversation. In a market where model releases are often judged against each other almost immediately, access and price can matter nearly as much as benchmark bragging rights.

For users, the practical implication is straightforward: if Carbon is cheaper and broadly available, more teams may be willing to test it in production workflows. That includes developers building products on the Gemini API, businesses using Gemini tools, and power users in the Gemini app. For Google, the upside is equally clear: more usage can mean more mindshare, more feedback, and more reasons for teams to keep building inside its ecosystem.

What is Antigravity’s ‘swarm’ feature?

Antigravity’s swarm feature is Google’s most interesting workflow experiment in this release. The Verge said Google is previewing a system that splits one prompt into coordinated agents. TestingCatalog AI News reported that the feature appears to use a bee-marked slash command to open a multi-agent dashboard with a board and work graph, allowing agents to take roles and coordinate through a shared message board.

In practical terms, swarm-style systems aim to divide a larger job into smaller pieces handled by separate AI agents. That can be useful for planning, coding, research, or any task where different sub-steps need to be tracked and handed off. The fact that Google is tying that concept to Carbon suggests the company wants the model not only to answer questions, but to operate as the engine for a more organized, multi-agent workflow.

What makes the swarm idea different from a normal chatbot?

It is different because the work is split and coordinated instead of being handled as a single stream of text. Rather than asking one model to do everything, a swarm design gives agents separate roles and a shared workspace. That could make complex tasks easier to manage, especially if Google can keep the output coherent.

TestingCatalog’s report adds a more visual detail: a dashboard with a board and work graph. While Google has not publicly detailed every part of the interface in the fact sheet, that reporting suggests the company is experimenting with a control layer that makes multi-agent work easier to oversee. If that holds up, Antigravity could become as much a project-management surface as an AI tool.

How cheap is Carbon compared with rivals?

Google is pricing Carbon aggressively, according to The Verge. The introductory price is $2 per million input tokens and $10 per million output tokens, with the promotional period ending at the end of November. The Verge also said that pricing comes in at roughly a fifth the cost of OpenAI GPT-6 Astra and about a quarter the cost of Anthropic Fable 5.2, at least by the paper comparisons cited in its report.

That pricing strategy matters because model choice is increasingly a budget decision, not just a quality decision. If a frontier model is substantially cheaper, teams can use it for more tasks, test it in more places, or route more work to it without blowing up costs. In a market that often rewards the newest model with a steep premium, Google is doing the opposite: trying to make its latest release feel accessible.

Background: how Google got here

Google’s Gemini line is its flagship AI model family, and Carbon is the latest step in that roadmap. The broader industry context is equally important: frontier releases have increasingly been judged by benchmark scores, coding performance, safety behavior, and token pricing. That has pushed companies to launch models that are not only more capable, but easier to justify economically.

AI model launches also tend to follow a familiar pattern: start with limited access, gather feedback from trusted users, then widen availability once the system is more battle-tested. The Verge reported that Carbon followed that path, beginning with a narrow rollout to cybersecurity professionals before moving to a broader public release across Google’s products and services.

The New Indian Express framed the model in security terms as well, saying Carbon improves vulnerability finding, exploitability assessment, and patch generation. The same report said Google saw stronger results in internal coding tests, including a more than twofold gain in secure-code success rate versus Argon, a nearly 80% drop in high-severity misalignment alerts in coding environments, and about half as many unsupported claims.

Those claims are important, but they should be read carefully. The outlets are describing different slices of the same release: The Verge focuses on rollout, positioning, and pricing; The New Indian Express emphasizes security and testing gains; TestingCatalog adds the interface details around Antigravity. Together they paint a picture of a release meant to do more than impress with raw capability.

How are outlets describing Carbon differently?

They are emphasizing different parts of the story. The Verge concentrates on public rollout, competitive positioning, and pricing. TestingCatalog focuses on the Antigravity interface and multi-agent workflow. The New Indian Express foregrounds security testing, coding performance, and internal benchmark-style results.

Those differences do not conflict; they just reflect different angles on the same launch. Taken together, they suggest Google is trying to make Carbon relevant to three audiences at once: developers who care about price and API access, security teams who need robust behavior, and advanced users who want agentic workflows rather than a simple chat box.

What the numbers say

Carbon’s reported figures point to a model that is being marketed as both capable and efficient. The New Indian Express cited an Artificial Analysis Intelligence Index score of 59, alongside security-related scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71% on CWE-bench. It also reported zero attack success rate on an indirect prompt-injection benchmark.

Those figures do not tell the whole story, but they help explain why Google is comfortable claiming Carbon is competitive with leading systems from Anthropic and OpenAI. In a field where models are compared by fractions of a point and cost-per-token differences can alter adoption, even modest gains can be strategically important.

By the numbers

• $2 per million input tokens during the introductory period.
• $10 per million output tokens during the introductory period.
• End of November is when the introductory pricing window closes.
• 59 is the reported Artificial Analysis Intelligence Index score.
• 88.2% is the reported SEC-bench Pro score.
• Zero is the reported attack success rate on the indirect prompt-injection benchmark.

Why it matters for developers and businesses

For developers, the biggest immediate effect is access. Carbon is now appearing through the Gemini API and AI Studio, which means builders can begin testing it in real products rather than waiting for a later release cycle. For businesses, the inclusion of Gemini for business broadens the chance that the model will be evaluated for workplace workflows, internal tools, and customer-facing systems.

The cheaper price also matters because it lowers the barrier to experimentation. Teams often choose models based on what they can afford to run repeatedly, not just what wins a benchmark. If Carbon is materially less expensive than competing frontier models, Google may get a wider set of use cases even before independent testing fully settles the performance debate.

For consumers and power users, the Gemini app for AI Ultra and Search AI Mode rollout could bring the model into everyday interactions rather than leaving it tucked inside developer tooling. That is where model branding becomes visible: not in a press release, but in the tools people touch day to day.

What happens next

The next questions are about availability, scope, and proof. The New Indian Express said wider access for developers, businesses, and consumers is planned later after a trusted-partner phase, while The Verge reported that Carbon is already moving into broader public circulation. The difference may simply reflect timing, but it also shows that rollout details can evolve quickly as Google expands access.

Watch for three things in the coming weeks. First, whether Carbon reaches lower-tier subscribers or free users. Second, whether Antigravity’s swarm feature remains exclusive to Antigravity or spreads into AI Studio or Google Cloud. Third, whether independent benchmark results support Google’s internal claims and the company’s confidence that Carbon can stand with leading Anthropic and OpenAI systems.

There is also the price clock. The introductory rates run through the end of November, which means users evaluating Carbon will need to watch what happens when those rates change. If the later pricing remains competitive, Google may keep the cost advantage that helped make the launch notable in the first place.

Timeline

The bigger competitive picture

Carbon arrives in a market where release cadence is almost part of the product. Google’s ability to roll out a cheaper frontier model, pair it with a novel multi-agent feature, and push it across multiple surfaces gives it a direct route back into the mainstream AI conversation. That is particularly significant after a stretch in which the spotlight has largely sat with Anthropic and OpenAI.

Still, the launch is not just about headlines. It is about whether Google can convert technical progress into visible momentum. Carbon’s rollout, its price, and the Antigravity swarm preview all point in the same direction: Google wants to look less like a follower and more like a company setting the terms of the next phase of AI products.

Whether it succeeds will depend on what users find when they actually put Carbon to work. The headline numbers are promising, but the real test will be how well the model performs once it leaves the controlled setting of an announcement and lands in the hands of developers, businesses, and consumers.