Google made Gemini 4 Carbon available to paying developers on Tuesday, positioning it as a step above Gemini 4 Argon for handling complex tasks in software engineering, cybersecurity, and scientific reasoning. Argon continues to serve as Google's standard recommendation for routine engineering and knowledge tasks, while Carbon targets more intensive workloads such as multi-package migrations, cross-subsystem debugging, and long-horizon agentic workflows.

Koray Kavukcuoglu, SVP of Google DeepMind and chief AI architect at Google, characterized Carbon in the launch announcement as the company's most capable model for intricate, multi-step reasoning. The release also brings /swarm, a preview capability within Google Antigravity that manages multiple coding agents from a single prompt.

Multi-agent coding arrives

Carbon can construct, validate, and repair full-stack applications, verifying both underlying code and the rendered user interface. Google indicates the model excels particularly at multi-package migrations and debugging across different subsystems.

What remains unclear is the mechanism by which /swarm distributes context among agents, resolves competing modifications, or determines which tasks warrant Carbon versus a less resource-intensive model. Google asserts that Carbon ranks first on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench, finishing assignments at under half the expense of the nearest competitor.

Pricing doubles November 30

Google has not furnished sufficient information to allow independent assessment of those cost claims. Carbon enters the market at $2 per million input tokens and $10 per million output tokens, equivalent to Argon's pricing, before increasing to $4 and $20 starting November 30. The question of whether Carbon's capabilities justify the elevated pricing, factoring in multi-agent coordination overhead and validation expenses, will likely only become apparent through real-world deployment.

Proving Signal's code

Among the more technically sophisticated demonstrations in Google's announcement is AlphaProof Code, a novel agentic framework designed for formal verification. Google's research team deployed Carbon alongside additional models to mathematically establish critical security characteristics of the cryptographic protocol underpinning Signal.

The researchers converted security requirements into mathematical theorems and leveraged AlphaProof Code agents to create millions of lines of Lean, a formal language for expressing and validating mathematical proofs. According to Google, the effort required weeks—substantially shorter than the years typically needed by specialized teams undertaking comparable verification work.

Google says the work took weeks, compared with the years that similar verification efforts have historically required from expert teams.

The announcement omits specifics about the complete range of verified properties, the extent of human involvement required, or whether Google intends to release the proof artifacts publicly.

Security that cuts both ways

Google initially rolled out Carbon's cybersecurity features to a restricted set of defenders via its Fairwind Program in late October. The company reports that Carbon leads performance on SEC-bench Pro, ExploitGym, and CWE-bench v1, benchmarks assessing vulnerability identification, exploitation, and remediation.

In Project Fortify, Google's autonomous scanning, categorization, and repair platform, Carbon achieved 100% scanning recall and 95% triage precision on an internal test. Google also notes a relative gain exceeding 25% in recall performance relative to Argon when evaluating open-source codebases.

Google's Cloud Vulnerability Research team employed Carbon to validate that two defects in a widely-used open-source initiative were exploitable, leading maintainers to mark them as critical and distribute fixes. Google indicates the flaws matched the severity of the recent PHP2Shell vulnerability, though the affected project remains unnamed pending completion of the disclosure timeline.

In a separate instance, Wiz leveraged Carbon through its Scan for Good initiative to identify zero-day vulnerabilities in two major SaaS platforms within under two hours. The impacted companies collaborated with researchers on patching.

Cybersecurity's double edge

The expanded exploitation capabilities introduce fresh misuse risks. Google indicates that Carbon's protective measures blocked 99% of attack paths tested by an adaptive red-teaming agent in controlled settings, bringing the per-attempt compromise likelihood down to 0.8%, versus 3.1% for Argon. The company also documents a 0% attack success rate on Gray Swan's indirect prompt-injection benchmark, though it concedes that more advanced attack vectors warrant continued investigation.

These results come from specific evaluation environments and don't establish that Carbon is resistant to all real-world attacks.

Getting access to Carbon

Gemini 4 Carbon is obtainable via the subscription-based Gemini API tier in Google AI Studio and Android Studio, alongside Google's commercial solutions. Gemini 4 Ultra subscribers can reach Carbon through Antigravity, the Gemini app, and AI Mode in Search. The /swarm orchestration capability remains in preview status.

Accessing Carbon's sophisticated cybersecurity tools still demands authorization via Fairwind. Google is additionally extending six months of complimentary access to Carbon through Antigravity to a cohort of selected researchers, with applications ongoing and fresh participants onboarded weekly. Google has released an updated Gemini 4 developer guide with migration and deployment guidance.