Google has announced its latest frontier model Gemini 4 Carbon, which the firm said outperforms the leading models from Anthropic and OpenAI and claims is a new industry leader in cybersecurity and long-horizon coding tasks.

The hyperscaler said Carbon was designed for complex agentic tasks and to build on the leap in coding performance introduced in Google’s earlier October release, Gemini 4 Argon.

In internal tests, Google added, its developers have let Carbon run for days on complex problems with no issue and have found the model capable at completing end-to-end cybersecurity and software development tasks.

The model is comfortably ahead of the competition in ExploitGym, a large-scale benchmark designed to evaluate AI agents on their ability to develop exploits, having scored 54.8 per cent.

This is significantly ahead of the 42.4 per cent scored by GPT-6 Astra and 34.5 per cent by Claude Opus 5.5.

Google added that Carbon is twice as capable at producing secure code for detected vulnerabilities than Argon, with its internal Chrome team stating that Cabon was over ten times more capable at generating patches than a competing frontier model.

In the benchmark CWE-bench v1, a rigorous test for how well AI models can patch real-world software vulnerabilities, Carbon scored 71 per cent, ahead of the 68 per cent achieved by Gemini 4 Argon and GPT-6 Astra.

Google said that in its internal testing, Carbon is far more trustworthy than prior models. The firm said it outputs around half of the unsubstantiated claims as Argon and was completely resistant to prompt injection in the Gray Swan Indirect Prompt Injection (IPI) Benchmark.

This compares to a 0.7 per cent success rate for Argon, one per cent for Claude Opus 5.5, 8.5 per cent for GPT-6 Astra, 51.8 per cent for Grok 4.6 and 52.7 per cent for Kimi K3.

The independent AI benchmarking platform Artificial Analysis ranks Carbon as the new leader for its fifth-generation intelligence index, which assesses AI model performance across nine benchmarks covering coding, science, and knowledge.

In the new index, Carbon scored 59 ahead of Claude Opus 5.5’s 58 and GPT-6 Astra’s 57.

Carbon will be available at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. In line with US government requirements on model pre-testing, it will first be rolled out to trusted cybersecurity partners and then to organisations enrolled in Google’s Fairwind Program.

This is Google’s frontier gated cybersecurity initiative for vetted companies, the equivalent of Anthropic’s Project Glasswing and OpenAI’s Daybreak. Approved organisations include Accenture, McKinsey & Company and Palo Alto Networks.

The model was announced just two weeks after the release of Google’s last major model, Gemini 4 Argon, marking a significant acceleration in the firm’s AI development cycle.

Since Demis Hassabis stepped down as the chief executive of Google DeepMind, the firm has shifted its focus more on direct competition in the LLM space, rather than longer-term research efforts.

Google DeepMind senior vice president Koray Kavukcuoglu, the former chief technology officer at Google DeepMind and chief AI architect at Google, has increased the pace of model development and deployment.

In September, he told the Information that his aim was to release frontier Gemini models “much earlier” than the end of year goal initially set by Google chief Sundar Pichai.

---