The iteration and updating of the AI industry is continuing to accelerate. Google DeepMind recently officially announced that its flagship large model with the strongest overall capabilities, Gemini4Carbon, is now officially launched, providing services to developers, enterprises, researchers, and ordinary consumers alike.
Targeting Highly Difficult, Complex Tasks: Strong Performance Across Multidimensional Benchmarks
As the core powerhouse in the Gemini4 family, Carbon focuses mainly on complex programming, cybersecurity, scientific reasoning, and multi-agent tasks that require long-duration execution. Previously, Google had opened the model to some cybersecurity defenders through the Fairwind program, and this launch further expands its scope of use.
In three authoritative benchmark tests—CursorBench4.0, Perplexity WANDR, and Zapier AutomationBench—Gemini4Carbon took first place in all of them, and its per-task cost was less than half that of the next-best-performing external competing model.
In terms of specific pricing and API access, Carbon's discounted API pricing is consistent with the previously released Argon: $2 per million input tokens and $10 per million output tokens, and from December 1 it will be adjusted to $4 and $20 respectively.
Disruptive Multi-Agent Collaboration: Building a Game Overnight and Reconstructing 3D Buildings from a Single Prompt
One of Gemini4Carbon's core killer features is its ability to efficiently coordinate multiple AI agents to jointly complete large, complex engineering projects. To this end, Google launched a preview of the /swarm feature in its AI programming platform Google Antigravity; users can enter a single instruction to organize multiple AI agents to divide up work and collaborate, while viewing progress in real time.
In the official demo, Carbon can play the roles of architecture design, task coordination, and code review, directing other faster and lower-cost sub-models to complete the actual development. For example, the team built an open-world 3D game from instructions alone, covering Blender 3D modeling, Three.js game development, sound effects production, and background music generated by Lyria; it even read 206 engineering drawings from 7 urban planning PDF files and, over 10 overnight runs, reconstructed an interactive 3D model of Google's Platform37 office building. In addition, it successfully restored a historical 3D exhibition and an antique piano with a 61-key mechanical structure.
Real-world use by third-party developers is equally impressive: software engineer Kit Langton had Carbon take over an unfinished Rust UI library development project before the end of the workday, and it was successfully delivered the next day; HubX tested 16 real bugs in a codebase with 14,000 tests, and Carbon ultimately achieved a 94% issue resolution rate without introducing new regressions.
Cybersecurity and Life Sciences: Finding Zero-Days in Two Hours and Accelerating Cryptography and Protein Research
In the cybersecurity field, Carbon leads across vulnerability discovery and remediation benchmarks such as SEC-bench Pro, ExploitGym, and CWE-bench. In Google's internal Project Fortify system, its vulnerability scanning recall reached 100% and its classification precision reached 95%. In real-world testing, it helped Wiz discover zero-day vulnerabilities in two large SaaS providers in less than two hours. At the same time, in simulated cyberattack tests, it successfully blocked 99% of attack paths launched by red-team agents.
In scientific research, Google announced three major breakthroughs:
Formal verification for cryptography: With the help of an agent framework called AlphaProof Code, Carbon transformed security goals into mathematical theorems and generated millions of lines of Lean proof code, completing a mathematical proof of key security properties in Signal's encryption protocol code within weeks.
Quality control for biological experiments: It can identify anomalies such as insufficient mixing in deep-well plates or residual instrument contamination from raw experimental data, improving experimental reproducibility.
Protein structure prediction: It assembled a multi-agent system to search massive databases for distant homologous proteins, increasing the average number of homologous sequences found by 40-fold on low-homology families in the CASP17 blind test.
