Google DeepMind has introduced Gemini 4 Carbon, a new artificial intelligence (AI) model designed to handle complex software engineering tasks, extended reasoning processes and advanced cybersecurity operations.
Announced on October 28, 2026, Gemini 4 Carbon is being positioned as Google’s strongest AI model so far, with particular improvements in identifying software vulnerabilities, developing security patches and completing complicated technical assignments that require several stages of work.
The announcement comes just two weeks after Google introduced Gemini 4 Argon, highlighting the company’s accelerating efforts to strengthen its position in the increasingly competitive global AI industry.
According to Google DeepMind Senior Vice President and Chief AI Architect Koray Kavukcuoglu, the new model represents another step towards developing AI systems capable of supporting more demanding professional and technical activities.
In its official announcement, Google explained that Carbon builds on the capabilities of Argon while delivering improvements across coding, agent-based workflows, scientific applications and cybersecurity defence.
Also Read: Can Africa Measure the Carbon Footprint of Online Education as Digital Learning Expands?
Gemini 4 Carbon Records Strong Performance in AI Benchmarks
One of the most significant aspects of Gemini 4 Carbon is its reported performance in independent AI evaluations and specialised cybersecurity benchmarks.
Google said the model achieved a score of 59 on the Artificial Analysis Intelligence Index version 5.0, placing it ahead of several competing frontier AI models, including Claude Opus 5.5.
The company also reported improvements in software development and automated problem-solving. These capabilities allow the model to work through complicated engineering assignments that may require sustained reasoning, multiple actions and repeated attempts before reaching a solution.
For web development, Gemini 4 Carbon recorded an improvement of 89 Elo points on WebDev Arena, reaching fourth position with a score of 1,768.
Its cybersecurity performance is another major selling point. Google reported that Carbon achieved 88.2 per cent on SEC-bench Pro, a benchmark designed to evaluate the identification of software vulnerabilities.
It also recorded 54.8 per cent on ExploitGym, which evaluates the ability to validate vulnerabilities through exploitation, and 71 per cent on CWE-bench, which measures how effectively AI systems can identify and fix weaknesses in software.
These results suggest that Google is increasingly interested in developing AI systems that can do more than generate code or answer technical questions. The company wants its models to identify problems, investigate their causes and help implement solutions.
Cybersecurity and Long-Running Tasks Take Centre Stage
Beyond its benchmark performance, Gemini 4 Carbon is designed to handle long-horizon tasks. This refers to assignments that require an AI system to maintain context, make decisions and complete several connected activities over an extended period.
Google said its internal teams have tested Carbon on complicated engineering problems, including identifying subtle bugs and implementing corrections across large software codebases.
According to the company, the model can continue working on demanding technical problems for days, navigating obstacles and adjusting its approach when earlier attempts fail.
Its cybersecurity capabilities are particularly relevant as organisations face increasingly sophisticated digital threats.
Carbon is designed to support different stages of cybersecurity defence, including discovering vulnerabilities, determining whether they can be exploited and developing patches to address the underlying weaknesses.
Google also reported improvements in the model’s security safeguards. These include stronger resistance to prompt injection attacks, a technique in which malicious instructions attempt to manipulate an AI system into ignoring its intended instructions.
The company said Carbon achieved a zero per cent attack success rate on the evaluated Gray Swan Indirect Prompt Injection benchmark.
However, Google is approaching the release carefully because powerful cybersecurity capabilities could also be misused if placed in the wrong hands.
Back Story: Google’s Growing Competition in the AI Industry
The introduction of Gemini 4 Carbon comes amid an intense competition between major AI developers, including Google DeepMind, OpenAI and Anthropic.
Companies in the sector are increasingly moving beyond conversational AI towards systems capable of performing complex professional assignments with limited human intervention.
Google’s introduction of Gemini 4 Argon in September 2026 established the foundation for its latest generation of frontier models. Carbon follows that development with a stronger emphasis on cybersecurity and sustained technical problem-solving.
As Reuters reported, the launch further intensifies competition in advanced coding and cybersecurity systems.
For now, Google is taking a phased approach to making Carbon available. The initial rollout is limited to selected cybersecurity professionals through its Fairwind Programme, with access expected to expand gradually to other trusted partners.
Google has indicated that the model will eventually become available to developers, businesses and consumers through its product platforms.
The introductory API pricing is set at $2 per million input tokens and $10 per million output tokens. Following the introductory period, the prices are expected to increase to $4 and $20 respectively.
Although wider availability is still pending, Gemini 4 Carbon represents another important development in the transition towards AI systems that can perform increasingly demanding technical and professional work.
Also Read: Google Gives Nigerian Students Free AI for One Year Across 27 African Countries
