Google DeepMind releases Gemini 4 Carbon for complex coding and security

A new frontier for engineering

Google DeepMind has introduced Gemini 4 Carbon, its latest frontier model released just two weeks after the Gemini 4 Argon model. Carbon has already secured the top spot on the Artificial Analysis Intelligence Index v5 with a score of 59 points. This marks a significant jump from the 50 points achieved by its predecessor.

The model is specifically engineered to handle much harder, long-horizon engineering problems. Unlike models designed for quick interactions, Carbon is built for persistent agentic workflows. This means it can work through dead ends on its own over multiple days to complete complex tasks.

Internal testing showed the model successfully managing intricate operations such as identifying the source of errors and resolving minor coding flaws within massive codebases. Google also reported improvements in scientific reasoning and multimodal understanding based on user feedback. It also saw a notable increase in performance for full-stack web development.

Security and coding benchmarks

Carbon shows significant gains in cybersecurity and programming capabilities. In the Gray Swan indirect prompt injection benchmark, the model reduced the attack success rate from 0.7 percent with Argon down to 0.0 percent. While this does not mean prompt injection is fully solved, the test appears to be saturated for the current model.

The model also performed well on the SEC-bench Pro, which evaluates the ability of an AI to find security flaws in actual software independently. It scored 88.2 percent and led all tested models in that specific category. Additionally, Carbon held a wide first-place lead in writing fixes on the CWE-Bench.

In coding, the model reached 1,768 points on the WebDev category of the Code Arena. This puts it in fourth place overall, trailing just two points behind Claude Sonnet 5.5. It also showed a net improvement of 16.2 percent in the Agent Arena, where it placed second across more than 21,900 real sessions.

Reasoning and research depth

Independent evaluations from Artificial Analysis show that Carbon leads the Intelligence Index v5 primarily due to its reasoning and knowledge retrieval. It took first place in four of the index's nine individual evaluations, particularly in scientific tasks. However, it still trails Claude Opus 5.5 in the Coding Agent Index.

The model also shows high proficiency in legal research and finance. On the Vals AI platform, Carbon recorded roughly 45 percent more cited sources and more frequent tool calls in legal research than Argon. It also leads both the Finance Index and the Legal Research benchmark on that same platform.

While it excels in depth, it can be slower for certain tasks. In the Finance and Tax Agent benchmark, tasks take more than twice as long as they did with Argon. In legal research, Carbon averages more than 50 minutes per task, generating 2.3 times as many reasoning tokens as its predecessor.

Limited access and rollout

Because of its advanced cybersecurity capabilities, Google is rolling out Carbon gradually. It is currently participating in the U.S. government's voluntary pre-testing process. This means the model will only be available to selected security experts starting this week.

Following the initial expert phase, Google plans to expand access through the Fairwind Program. This program will provide the model to other trusted partners before it reaches the general public. Google intends to eventually open the model to all consumers and businesses through its standard product interfaces.

The pricing for Gemini 4 Carbon is set at the same introductory rate as Argon. Users will pay $2 per million input tokens and $10 per million output tokens. Cached input tokens will receive a 95 percent discount off the regular input price during this period.