Google released Gemini 4 Carbon on October 28, 2026, marking the latest update to its frontier AI models just two weeks after the debut of the Argon variant. According to a post by Sundar Pichai on X, the new model is designed to improve performance in complex, long-horizon tasks and cybersecurity applications. As Google said, the company has adopted a rapid release cycle, gradually introducing new models within the Gemini 4 series. Koray Kavukcuoglu, senior vice president of Google DeepMind, who assumed leadership of the research lab following the departure of Demis Hassabis, described the release as “the most powerful Google model to date.”
Gemini 4 Carbon Tops Intelligence Index Leaderboard
The Artificial Analysis Intelligence Index v5.0 placed Gemini 4 Carbon at the top of its leaderboard with a score of 59 points. This score narrowly edges out the Claude Opus 5.5 model, which holds 58 points, and the GPT-6 Astra model from OpenAI, which stands at 57 points. The index represents a summary of nine tests, approximately 60% of which consist of private tasks to prevent models from being pre-trained on the evaluation material.
I did say we'd be releasing our models periodically…introducing Gemini 4 Carbon, our next frontier model. Following Argon's release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of @ArtificialAnlys… pic.twitter.com/GN9LqZ7Zcf
Internal data from Google suggests a broader lead in specific technical domains. In a series of 17 internal benchmarks, Google claims Carbon outperformed other models in 15 categories, including significant improvements in code vulnerability detection and patching. Carbon achieved 88.2% in the SEC-bench Pro test for vulnerability discovery, 54.8% in ExploitGym for vulnerability exploitation, and 71.0% in the CWE-bench test for patching. Carbon demonstrated a 96.2% success rate in the GraphWalks test for handling contexts of up to one million tokens, compared to 71.8% for Astra and 66.8% for Opus. As Artificial Analysis reports, while Carbon leads in overall reasoning, Claude Code scores 62 points in agent programming with the Opus 5.5 model, while Carbon scored 59 points in the Antigravity CLI environment.
Google Targets Hallucinations and Security Bypasses in Carbon
Google has prioritized safety and accuracy in the Carbon release, specifically targeting the reduction of “hallucinations” and unauthorized command execution. The company reports that Carbon produces roughly half as many unsubstantiated claims as its predecessor, Argon.
Introducing Gemini 4 Carbon, our latest frontier model. Carbon delivers improved performance across domains — ranking #1 on v5.0 of the @ArtificialAnlys Intelligence Index and sets a new state-of-the-art in cybersecurity and long-horizon capabilities. It’s rolling out starting… pic.twitter.com/oQrHzxvMSo
Testing data shared by the company indicates that the model showed zero successful “jailbreak” attempts in Terminal-Bench 4.0, with only 0.4% of attempts resulting in a successful bypass. In tests conducted by Gray Swan regarding indirect prompt injection, the model reportedly resisted all attempted exploits. The model incorporates monitoring of the inference process and additional alignment work to ensure it functions according to expectations.
Carbon Access Restricted to Cybersecurity Researchers and Partners
Access to Gemini 4 Carbon is currently restricted to a select group of cybersecurity researchers and partners via Google’s “Fairwind” program. A public release date for general users or enterprise customers has not been established.
When the model eventually reaches the API, Google has committed to a pricing model identical to the Argon release. Costs are set at $2 per million input tokens and $10 per million output tokens, with cached input tokens offered at a 95% discount. As Vietnam.vn reported, this pricing strategy positions Gemini 4 Carbon as a lower-cost alternative to competitors like GPT-6 Astra, which is priced at $10 per million input tokens and $50 per million output tokens, and Claude Opus 5.5, which costs $4 per million input tokens and $20 per million output tokens.
Future Integration and Data Usage
Google continues to explore diverse applications for the Gemini 4 architecture, including research into robotics, such as training humanoid systems for physical tasks like cornhole.
Answers to Common Questions About Gemini 4 Carbon
How does Gemini 4 Carbon differ from Gemini 4 Argon?
Carbon is an iterative advancement rather than a new generation. While it shares the same underlying security architecture as Argon, it demonstrates higher reasoning capabilities and lower rates of factual errors in technical tasks.
When will the public be able to use Gemini 4 Carbon?
Google has not provided a definitive timeline for a public launch. The model is currently in a testing phase, limited to a small group of cybersecurity experts, following a similar pattern to the initial rollout of the Argon model.
Announcing Gemini 4 Carbon, our strongest model with state-of-the-art cybersecurity and long horizon capabilities. Carbon represents a step up across most benchmarks, delivering strong frontier performance in autonomous coding and cyber defense. We’re first making Carbon… pic.twitter.com/jgrKXp3RYl
What is the significance of the “Fairwind” program? Fairwind is the internal designation for the program through which Google grants early access to its frontier models to trusted partners and security professionals.