Figure 1: Official Key Visual for Gemini 4 Carbon (Source: Google Official)
On November 11, Google officially announced the launch of Gemini 4 Carbon, its most powerful artificial intelligence model to date. Developed by the Google DeepMind team, this model is designed for long-cycle agent workflows involving complex knowledge work, software engineering, cybersecurity, and scientific reasoning. It is now available to paying developers and enterprise users via the Gemini API in Google AI Studio and Gemini for Business.
Positioning: Flagship Model for Tasks Pushing the Limits of Intelligence
Koray Kavukcuoglu, Chief AI Architect at Google and Senior Vice President of Google DeepMind, stated in the release article that Carbon is positioned to tackle tasks that challenge the limits of model intelligence, such as multi-step coding and debugging, and open-ended scientific research. Compared to the previously released Gemini 4 Argon, which still handles the majority of routine software engineering and knowledge work, Carbon performs deep exploration, rigorous reasoning, and extended verification cycles by consuming more tokens. Reportedly, Carbon was made available early to some cybersecurity experts through the "Fairwind Program" two weeks before its official launch.
Ranked First in Multiple Benchmarks, Launch Price Matches Argon
According to data released by Google, Carbon ranked first in three independent benchmarks: CursorBench 4.0, Perplexity WANDR, and Zapier AutomationBench. Additionally, its cost per task is more than twice as low as that of the next best comparable model.
To address the issue that existing benchmarks struggle to measure an agent's ability to handle dense files such as charts, blueprints, and tables, Google partnered with Artificial Analysis to launch the Multimodal Knowledge Agent Benchmark (MKA). Carbon ranked second in this test, with an average cost per task more than six times lower than the top-ranked model. In the financial sector, Carbon ranked first among 23 frontier model configurations in Balyasny Asset Management's General Financial Intelligence Benchmark; Hebbia's end-to-end financial agent evaluation showed it performed on par with the best tested models.
Regarding pricing, Carbon's initial introductory price is the same as Argon: $2 per million input tokens and $10 per million output tokens. The promotional period will end on November 30, 2026, after which prices will revert to $4 per million input tokens and $20 per million output tokens.
/swarm Preview Launches, Generating Collaborative Agent Teams from a Single Prompt
Coinciding with the new model launch, Google introduced the /swarm preview in its agent development platform, Google Antigravity. This feature converts a single prompt into a collaboratively tracking agent team: Carbon handles roles requiring deep reasoning, such as architecture design, coordination, and auditing, while working alongside lightweight models and adversarial validators in parallel isolated workspaces.
Engineering Development: Multi-Agent Collaboration to Build Runnable Applications
In engineering scenarios, Google states that Carbon supports multi-package migration, cross-subsystem troubleshooting, and architecture refactoring. It can run "build and verify" loops, automatically recover from errors, and simultaneously check code and rendered UIs, making it suitable for web applications, 3D game development, and creative programming. Several early-testing enterprises provided positive feedback: HubX reported a 94% fix rate with zero regressions on 16 real defects within its codebase of 14,000 test cases; OpenCode engineers stated that Carbon built a Rust UI library from scratch and rewrote a desktop application overnight.
Cybersecurity: Covering Vulnerability Discovery, Exploitation, and Patching
The newly released Carbon for cyber is described by Google as its strongest cybersecurity model to date, leading in benchmarks such as SEC-bench Pro, ExploitGym, and CWE-bench v1.
Google's disclosed internal practices show that in its autonomous vulnerability system, Project Fortify, Carbon achieved a 100% scan recall rate and 95% classification accuracy, improving vulnerability detection capability by over 25% compared to Argon in open-source code repository tests. The Google Cloud Security Vulnerability Research Team used this model to confirm that two bugs in a popular open-source software were exploitable vulnerabilities, leading to their reclassification as high-risk. Cloud security vendor Wiz, in its "Scan for Good" project, used Carbon to discover 0-day vulnerabilities in two large SaaS vendors within two hours, a process that typically takes weeks using traditional methods. Partners such as SentinelOne, Cisco, and Chainguard have also confirmed its performance in vulnerability chain analysis and long-context scenarios in official materials.
Scientific Reasoning: From Encryption Protocol Verification to Protein Research
In the scientific research field, Carbon has been used by drug toxicity research company Axiom Bio to discover new toxicity mechanisms. Google's internal research team also demonstrated three applications: first, using an agent framework called AlphaProof Code to complete formal mathematical proofs of key security properties in the underlying code of the Signal encryption protocol within weeks, a task that previously required expert teams years; second, identifying wet lab biases and improving chromatographic analysis methods in enzyme assay development; and third, increasing the average number of retrieved homologs by 40 times in seven low-homology protein families from the CASP17 blind test set, providing richer upstream information for protein structure prediction.
Furthermore, Google's previously announced plan to provide free six-month access to Antigravity for 1,000 scientists has now been expanded to include the Carbon model. The application channel remains open with rolling weekly reviews.
Security Assurance: Prompt Injection Attack Success Rate Drops to Zero for the First Time
Addressing potential misuse risks associated with cybersecurity capabilities, Google simultaneously released a set of security data: in tests simulating "adaptive red team agents," Carbon blocked 99% of attack paths, with a single-attempt breakthrough rate of 0.8%, lower than Argon's 3.1%. In the Gray Swan test, the success rate of automated cyberattack queries dropped to 2.17%, compared to Argon's 9.86%. External organization Trajectory Labs found no new universal jailbreak methods during tests lasting 132 hours with over 13,000 requests.
In terms of alignment performance, Carbon achieved a 92% pass rate in the Agent Honesty Benchmark containing 375 tasks, compared to Argon's 70%. On the Gray Swan IPI benchmark, the success rate of indirect prompt injection attacks dropped to 0% for the first time. Google also acknowledged that there is still room for continuous improvement in specific scenarios when facing complex and sophisticated attacks.
Figure 2: Comparison of model alignment metrics across the full lifecycle; lower values are better (Source: Google Official)

