Originally published on The Daily Flare.
Google announced Gemini 4 Carbon, describing it as its strongest frontier model yet and targeting cybersecurity defense alongside extended coding and engineering tasks.
The company says the model leads several cybersecurity benchmarks covering vulnerability discovery, exploitation and patching. Google is beginning access with selected cyber defenders and plans to expand it gradually to other trusted partners.
Benchmark results span the defense process
In its announcement, Google said Gemini 4 Carbon scored 88.2% on SEC-bench Pro, which tests vulnerability discovery; 54.8% on ExploitGym, focused on exploitation; and 71.0% on CWE-bench, which assesses patching. Google described the results as state-of-the-art. The company said these external benchmark evaluations were run without guardrails, a qualification that matters when interpreting scores for a model intended for security work.
The model also ranked first on the recently updated v5.0 Artificial Analysis Intelligence Index, with a score of 59, according to Google. The company reported gains in coding and agentic tasks, where models carry out sequences of actions toward a goal. It did not provide further details about the index methodology in its announcement.
Google’s internal tests found more than a twofold increase over Gemini 4 Argon in success at generating secure code. The company also said its Chrome team saw Carbon produce high-quality patches at more than 10 times the rate of another large frontier model from an unnamed lab. Google did not identify that model or provide a comparison methodology in the announcement.
Wiz, Google’s cloud security platform, also saw Carbon lead on web penetration-testing benchmarks, the company said. Google linked that capability to Wiz’s Scan for Good initiative. The announcement did not provide a score for those tests.
Long-running engineering work and staged access
Google says Carbon is designed for extended, multi-step tasks rather than only short exchanges. Its teams have used the model internally on engineering problems that can take days, including tracking down subtle bugs and developing fixes across large codebases and low-level systems. The company said the model can continue working through dead ends, but did not describe a public test or independent evaluation of those internal examples.
The release follows Gemini 4 Argon, which Google says launched two weeks earlier. The company also highlighted improvements in scientific tasks, full-stack web development and multimodal understanding. On WebDev Arena, a ranking based on comparative evaluations of web-development outputs, Google said Carbon gained 89 Elo points to reach fourth place with a score of 1,768.
Google is taking a staged approach to availability. Starting this week, a subset of cyber defenders will receive access, with additional trusted partners expected to join gradually through the company’s Fairwind Program. Google said it is participating in the U.S. government’s voluntary pre-release model-testing process and will continue to gather feedback and refine safeguards before broader access.
The company has not opened Carbon to general developers, businesses or consumers. It said it intends to offer the model through its product services as soon as possible, after further work on guardrails. That leaves the initial rollout limited to selected users while Google tests the system in security-focused settings.
Pricing announced ahead of wider release
Google set the planned introductory price at $2 per million input tokens and $10 per million output tokens, matching Argon’s introductory rates. Cached input tokens will cost 95% less than the standard input price. These are the announced prices for when Carbon becomes available through Google’s product surfaces, not a sign that broad public access has begun.
The release puts the emphasis on defensive uses, including finding and fixing software weaknesses, while the company’s own announcement also notes performance in exploitation testing. For now, access remains restricted as Google gathers partner feedback and works on safeguards before deciding on broader availability.
Top comments (0)