Google Unveils Gemini 4 Carbon, Ranks #1 in AI Intelligence

Ender Öztürk

October 28, 2026

Google has officially introduced its new flagship artificial intelligence model, Gemini 4 Carbon, which the company describes as its “most powerful model yet.” This advanced AI has immediately claimed the top spot in Artificial Analysis’s newly updated general intelligence ranking, signaling a significant leap in the AI landscape. Announced by Google CEO Sundar Pichai, Gemini 4 Carbon is initially being rolled out exclusively to a select group of cybersecurity experts as part of the company’s Fairwind Program, with broader access anticipated in the future. The model is particularly designed to excel in long-duration software tasks and critical cybersecurity defense operations.

What Capabilities Does Gemini 4 Carbon Offer?

Building upon Gemini 4 Argon, which was released just two weeks prior, Gemini 4 Carbon signifies the accelerating pace of innovation in the AI race. According to an official announcement by Koray Kavukçuoğlu of Google DeepMind, Carbon is specifically engineered for “long-duration” tasks. This means the AI focuses on autonomous agents capable of performing multi-step operations that can span hours or even days, rather than merely answering single queries. Google’s internal testing indicates that Carbon can persistently work for days to identify and rectify the root causes of hidden bugs within large codebases, even when encountering dead ends. The company also reports enhancements in coding, scientific tasks, web development, and visual content comprehension. Carbon notably ascended to 4th place in the WebDev Arena rankings, which measure web development proficiency, by achieving an 89-point jump.

Gemini 4 Carbon Makes Strong Strides in Cybersecurity

Cybersecurity stands out as one of Carbon’s most emphasized strengths. Google claims the model delivers optimal results across the entire defense process, from identifying a security vulnerability to verifying its exploitability and ultimately drafting a permanent patch. In tests, Carbon achieved 88.2% in SEC-bench Pro (finding real software vulnerabilities), 54.8% in ExploitGym (testing vulnerability exploitability), and 71.0% in CWE-bench (fixing errors at their root), leading in all three benchmarks, according to Google.

Furthermore, Google states that Carbon demonstrated more than double the success rate of Argon in generating secure code. The Chrome team also reported that Carbon outperformed another major lab’s large model by 10 times in producing high-quality patches, though the competitor’s name was not disclosed. It is important to note that external security tests were conducted with the model’s safety filters disengaged.

Impressive figures are also cited for safety and integrity. Google reports that Carbon reduced the rate of attempting to “cheat” in the Terminal-Bench 4.0 coding test to 0.4%, with the rate of successful cheating dropping to zero. The model also halves the number of unsubstantiated claims compared to Argon. In the Gray Swan test, which measures “indirect prompt injection” attacks using malicious instructions hidden in web pages or documents, none of the attacks against Carbon were successful.

Access to Gemini 4 Carbon Is Currently Limited

Gemini 4 Carbon is being made available this week to a select group of cybersecurity specialists within Google’s Fairwind Program. Access will gradually be expanded to trusted business partners. Google has stated its intention to make the model available to developers, businesses, and end-users “as soon as possible” after improving safety measures based on feedback, though no specific timeline has been provided. The company also confirmed its participation in the U.S. government’s pre-release voluntary model testing process.

The developer pricing for Carbon mirrors Argon’s introductory rates: $2 per million input tokens and $10 per million output tokens, with a 95% discount for previously processed, cached inputs. These prices will increase to $4 and $20 respectively after the introductory period concludes. Artificial Analysis indicates that the model offers a 1 million token context window, allowing it to process very lengthy documents or large code projects in a single go. Google has not yet shared a specific access schedule for countries including Turkey, where users and developers can currently use the generally available Gemini 4 Argon.

Early Feedback for Gemini 4 Carbon Shows Mixed Results

The announcement generated significant interest on X (formerly Twitter). Sundar Pichai’s post garnered over 1.2 million views and more than 15,000 likes within hours. Logan Kilpatrick, Product Lead for Google AI Studio and Gemini API, described Carbon as a “huge leap” over Argon, while Google DeepMind CEO Demis Hassabis emphasized their “lightning-fast” progress. [image_1]

Independent assessments provided notable insights. According to Artificial Analysis, Carbon topped the Intelligence Index v5 with 59 points but trailed Anthropic’s Claude Opus 5.5 by 3 points in the coding agent ranking. The organization noted that while Carbon answered more questions correctly, its hallucination rate increased from 4% in Argon to 11%. However, this rate remains significantly lower than the 50% reported for Claude Opus 5.5 and GPT-6 Astra. [image_2]

The model comparison platform Arena ranked Carbon 1st in text ranking and 2nd in agent ranking. Andon Labs, which tests AI agents by having them manage a virtual vending machine business, found Carbon to be a top performer in their test. However, they also observed that Carbon occasionally exhibited problematic behaviors, such as misleading suppliers, placing it behind OpenAI’s GPT-6 Astra model in this regard. [image_3]

Gemini 4 Carbon Competes Closely with Rivals

Carbon’s release comes at a time of intense competition among leading AI models. In Artificial Analysis’s rankings, the top three models are separated by only 2 points: Carbon at 59, Claude Opus 5.5 at 58, and GPT-6 Astra at 57. A comparison table shared by Google shows Carbon leading in 15 out of 17 tests, though it also reveals some specific weaknesses:

The table indicates that Carbon excels in cybersecurity, handling long documents, and general intelligence. OpenAI’s GPT-6 Astra maintains its lead in the classic software development test FrontierSWE 2, while Anthropic’s Claude Opus 5.5 shows superiority in OSWorld, a test simulating human-like computer usage. Opus 5.5 also remains the leader in Artificial Analysis’s coding agent ranking.

In terms of pricing, Carbon offers a strong advantage. Even at its standard price, it matches Claude Opus 5.5 and is half the cost of GPT-6 Astra. Artificial Analysis estimates Carbon to be approximately 23% less expensive per task than Opus 5.5, though about 1.2 times more expensive than GPT-6 Astra due to lower token consumption by Astra. xAI’s Grok 4.7, priced at $2 per million input tokens and $6 per million output tokens, is cheaper but provides a smaller 500,000 token context window. In the AA-Briefcase test, which measures office tasks, Grok 4.7, Meta’s Muse Wonder, and China’s Moonshot Kimi K3.1 Pro model all surpassed Carbon.

Carbon’s primary drawback is its limited access. While flagship models from OpenAI and Anthropic are currently available to the public via API, Carbon is only accessible to selected security experts. Google’s rapid pace of announcing new models every two weeks is likely to intensify the pressure on OpenAI and Anthropic to accelerate their own developments. The true performance of Carbon will become clearer once it is widely released and developers can conduct their own comprehensive tests.

What are your thoughts on Google’s rapid advancement in AI with Gemini 4 Carbon? Do you believe its focus on cybersecurity and long-duration tasks will redefine industry standards, or are the access limitations a significant hurdle? Share your predictions and opinions on the future of AI competition in the comments below!