Google Launches Gemini 4 Carbon Model and Agent Swarm Feature

Google DeepMind has announced the public release of Gemini 4 Carbon, described as its strongest artificial intelligence model to date [1][2]. The model is rolling out following a limited release two weeks prior to selected cybersecurity defenders [1][2].

Capabilities and Architecture

Google DeepMind states that Gemini 4 Carbon pushes capabilities further than its predecessor, Gemini 4 Argon, which was launched a month prior [2]. While Argon is positioned as suitable for most standard software engineering and knowledge work, Carbon is designed for intensive tasks that test model intelligence limits, including multi-step coding, debugging, and open-ended scientific research [1][2]. The model utilizes more tokens than Argon to facilitate deep reasoning, thorough exploration, and extended verification loops [1].

Across multiple independent benchmarks, Carbon has demonstrated strong performance in real workflows. It ranks first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench [1]. In legal, financial, and enterprise operations, the model handles visual understanding, reasoning, and document synthesis, following strict schemas to generate spreadsheets, presentations, and reports [1]. Google partnered with Artificial Analysis to introduce the Multimodal Knowledge Agent benchmark, where Carbon placed second [1]. Furthermore, Carbon matched top-performing models in pre-release evaluations of deep financial research on Hebbia's E2E Financial Agentic evaluation and Balyasny Asset Management’s General Financial Intelligence benchmark [1].

The Antigravity Swarm Feature

Alongside the model release, Google is previewing a new /swarm feature within its Google Antigravity platform [1][2]. This feature allows a single user prompt to generate a coordinated team of AI agents that can be tracked in real time [1][2]. Developers can assign Carbon to complex architectural design, coordination, and auditing roles, while pairing it with lighter models and adversarial verifiers across parallel isolated workspaces [1]. Examples demonstrated by Google include reconstructing historical buildings, generating 3D games, and building multi-scale textile visualizers [1].

Cybersecurity and Scientific Applications

Carbon delivers advanced cybersecurity intelligence for protecting critical software, leading on benchmarks for finding, exploiting, and patching vulnerabilities when accessed through the Fairwind Program [1]. Google reported success in internal testing via Project Fortify, open-source repository scanning, and bug discovery alongside external organizations such as Wiz [1]. In the life sciences, Axiom Bio identified carbon as a leading model for translational science, uncovering novel mechanisms of drug toxicity [1].

Pricing and Availability

Gemini 4 Carbon is available via the API in Google AI Studio and Gemini for business for paying developers and enterprises [1][2]. It is also rolling out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search, and Google Antigravity [1][2]. The model launches at an introductory price of $2 per million input tokens and $10 per million output tokens, which is scheduled to double at the end of November [1][2].

The release follows Google's rollout of the Gemini 4 Argon model [1][2]. For several months, the forefront of frontier AI development was widely perceived as a two-horse race between OpenAI and Anthropic, making Google's rapid succession of model releases a notable shift in the competitive landscape [2].

The introduction of Gemini 4 Carbon and the Antigravity /swarm feature expands the operational capacity of AI agents in enterprise, cybersecurity, and scientific research environments. Competitive token pricing impacts the economics of deploying frontier AI models across Indian and global markets.

The release of lower-cost, highly capable frontier models could pressure competitors like OpenAI and Anthropic on pricing and enterprise adoption. The multi-agent swarm architecture may accelerate autonomous software development and vulnerability research workflows.

It remains unconfirmed when or if Carbon will be made available to lower-tier subscribers or free users [2]. Benchmark results can be selectively chosen by companies to present models favorably [2].