Google Launches Gemini 4 Carbon, Claims AI Model Verified Signal's Encryption in Weeks

Google made Gemini 4 Carbon available to paying developers on Tuesday, establishing it as the company's most advanced frontier model for handling intricate software engineering, cybersecurity, and scientific reasoning challenges. Gemini 4 Argon continues to serve as Google's standard recommendation for routine engineering and knowledge tasks, whereas Carbon targets more intensive applications such as multi-package migrations, cross-subsystem debugging, and extended agentic workflows.

Koray Kavukcuoglu, senior vice president of Google DeepMind and the company's chief AI architect, characterized Carbon in the announcement as Google's most capable model for intricate, sequential work. The rollout also introduces /swarm, a preview capability within Google Antigravity that orchestrates multiple coding agents from a single instruction.

Multi-agent coding arrives

Carbon possesses the ability to construct, validate, and troubleshoot complete applications, verifying both underlying code and the resulting user interface. According to Google, the model demonstrates exceptional performance on multi-package migrations and debugging spanning multiple subsystems.

Google has not disclosed specifics regarding how /swarm distributes information across agents, manages competing modifications, or determines when Carbon is necessary versus employing a less resource-intensive alternative.

Google asserts that Carbon achieves top rankings on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench, finishing assignments at under half the expense of competing leading models.

Pricing doubles November 30

The company has not supplied sufficient information to allow independent confirmation of those cost assertions. Carbon enters the market at $2 per million input tokens and $10 per million output tokens, equivalent to Argon's current pricing, before increasing to $4 and $20 starting November 30.

The question of whether Carbon's capabilities justify the elevated pricing, factoring in the expenses of multi-agent management and validation, will depend on real-world deployment outcomes.

Proving Signal's code

Among the more technically sophisticated demonstrations in Google's announcement is AlphaProof Code, a fresh agentic system designed for formal verification.

Google's research team employed Carbon alongside additional models to formally establish critical security attributes of the cryptographic protocol underlying Signal. The researchers converted security criteria into mathematical formulations and leveraged AlphaProof Code agents to produce millions of lines of Lean, a language for articulating and validating mathematical proofs.

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

The announcement does not clarify the complete range of verified characteristics, the degree of manual involvement needed, or whether Google intends to release the proof materials to the public.

Security that cuts both ways

Google initially made Carbon's security capabilities accessible to a restricted group of defenders via its Fairwind Program beginning in late October.

According to Google, Carbon leads performance on SEC-bench Pro, ExploitGym, and CWE-bench v1, evaluation frameworks assessing vulnerability identification, exploitation, and remediation.

Within Project Fortify, Google's system for independent scanning, categorization, and fixing of vulnerabilities, Carbon achieved complete scanning recall and 95% categorization accuracy on an internal evaluation. Google also notes a gain exceeding 25% relative to Argon's recall performance when examining open-source codebases.

Google's Cloud Vulnerability Research division employed Carbon to establish that two defects in a widely-used open-source initiative could be leveraged, leading maintainers to designate them as critical and issue corrections. Google characterizes the vulnerabilities as equivalent in severity to the recent PHP2Shell vulnerability but has withheld the project's identity pending completion of the disclosure timeline.

In a separate instance, Wiz leveraged Carbon via its Scan for Good initiative to discover zero-day vulnerabilities in two major SaaS platforms in fewer than two hours. The impacted organizations collaborated with researchers on fixes.

Cybersecurity's double edge

The heightened attack capabilities introduce concerns about potential misuse. Google indicates that Carbon's protective mechanisms prevented 99% of attack paths executed by an adaptive red-teaming agent in test scenarios, bringing the per-attempt compromise probability down to 0.8%, in contrast to 3.1% for Argon.

Google also documents a 0% attack success rate on Gray Swan's indirect prompt-injection assessment, while recognizing that more advanced attack methods remain a domain requiring additional work.

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 accessible in Google AI Studio and Android Studio, alongside Google's business solutions.

Subscribers to Google AI Ultra 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 security capabilities still demands clearance through Fairwind. Google is also providing selected researchers with six months of complimentary access to Carbon via Antigravity, with applications being accepted and fresh researchers onboarded weekly.

Google has released an updated Gemini 4 developer guide addressing migration and deployment procedures.