Google has officially launched global access to its flagship AI model, Gemini 4 Carbon, making it available to developers, enterprise clients, research teams, and regular users. The company positions Carbon as a tool 'for tasks at the edge of AI capabilities'—in contrast to Gemini 4 Argon, which remains the optimal solution for everyday analytics and standard programming. Carbon's key differentiator is that the model consumes significantly more compute resources for deep reasoning, multi-stage hypothesis testing, and extended autonomous task execution without human intervention. This architecture enables the neural network to handle complex tasks that previously required whole teams of specialists.
Development Autonomy: 94% Successful Solutions Without Side Bugs
Internal tests conducted by Google's team prior to the official release recorded impressive metrics of model autonomy. In a standardized benchmark for bug fixing across a codebase consisting of 14,000 tests, Carbon achieved a 94% success rate in resolving issues without introducing regressions or new bugs. Developers emphasize that the model does not just generate a patch; it conducts multi-step context analysis, evaluates alternative solutions, and verifies the result before output. This fundamentally sets Carbon apart from lighter models, which in similar scenarios often introduce new defects while fixing old ones.
Swarm: A Single Prompt Becomes a Team of AI Agents
The release's main technological breakthrough is the Swarm tool within the Google Antigravity environment. The feature allows a user's single text prompt to be transformed into a coordinated group of AI agents operating in real time as a cohesive team. In this hierarchy, Carbon handles tasks requiring the deepest level of analysis—architecture design, agent coordination, and final auditing—while simpler subtasks are delegated to fast, lightweight models operating in parallel isolated spaces. During public demonstrations, the agent 'swarm' built an interactive 3D exhibition of the Eiffel Tower from 1887 photographs based on a single prompt, reproduced a Google office building from 206 PDF blueprints, and created a fully functional browser-based 3D game.
Cybersecurity: From Zero-Days to Automated Patching
As part of the special Fairwind program, the model was provided to cybersecurity specialists, and Carbon demonstrated top-tier performance in vulnerability discovery, exploitation, and automated patching. Within Google's autonomous Project Fortify system, the neural network detected all errors during code scanning, and the accuracy of identifying hidden logical vulnerabilities in open-source software increased by 25%. The most high-profile result came from the Wiz Scan for Good initiative: within two hours, Carbon discovered a zero-day vulnerability in two major SaaS providers, compressing a process that usually takes weeks of manual research into a single workday.
Science, Cryptography, and Finance: Results Beyond IT
Carbon delivered strong results across fundamental disciplines including physics, mathematics, and biology. Axiom Bio is already transitioning its production processes to the new model after the AI identified previously unknown toxicity mechanisms in drug compounds. The AlphaProof Code utility, utilizing Carbon, mathematically proved the security of the Signal cryptographic protocol in the Lean programming language within a matter of weeks—a task that typically takes human experts years of work. In the financial sector, the neural network topped the Balyasny Asset Management benchmark, outperforming 23 configurations of other models across aggregated metrics.
Pricing: Twice as Cheap per Task Than Competitors
Google has set initial pricing at $2 per 1 million input tokens and $10 per 1 million output tokens. According to the company, this is more than twice as cheap as the closest external competitors in terms of task execution cost and six times cheaper in multimodal benchmarks. Thus, access to a model capable of autonomously finding zero-day vulnerabilities and conducting formal proofs has become economically viable for a broad range of developers and research groups.
