Google DeepMind announces Gemini 4 Carbon, aimed at long engineering tasks and cyber defense

Google DeepMind has announced Gemini 4 Carbon, which it calls its strongest model. The company says Carbon has state-of-the-art cybersecurity and long-horizon capabilities. Carbon arrives two weeks after Gemini 4 Argon, the model it builds on.

The release is limited. Starting this week, Carbon goes only to a subset of cyber defenders. Access will then widen to trusted partners through Google's Fairwind Program. Developers, businesses and consumers will get it later. Google gives no date and says only "as soon as possible". The company describes its rollout as "deliberate" and says it is taking part in the U.S. government's voluntary process for pre-release model testing.

What Carbon is said to do that earlier models did not

Google's central claim is about duration. According to the announcement, Carbon is "built for sustained agentic workflows". Internally, Google says, its teams have seen Carbon "run for days to persist through dead ends" on complex, multi-step engineering tasks. The examples it gives are root-causing and patching subtle bugs across large codebases and low-level systems.

The second claim is about security work. Google says Carbon covers the whole defensive cycle: finding a vulnerability, validating it and fixing it. Its reported scores, evaluated without guardrails, are:

• 88.2% on SEC-bench Pro, for finding vulnerabilities
• 54.8% on ExploitGym, for exploitation
• 71.0% on CWE-bench, for patching

Google says these scores give Carbon "a firm lead" on all three benchmarks. It did not publish the full comparison table as text in the announcement, so the margins over other models are not stated there.

Google also cites results from its own teams. On internal benchmarks, it says, Carbon has more than twice Argon's success rate at generating secure code. According to the announcement, the Chrome team reported that Carbon did more than 10 times better than "another lab's large frontier model" at generating high-quality patches. Google did not name that model. Wiz, the cloud security company Google owns, says Carbon leads on its benchmarks for web penetration testing. That capability is central to Wiz's Scan for Good initiative.

Broader benchmarks and price

Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. On WebDev Arena, a leaderboard for web development, it moved up 89 Elo points to 1768. That puts it fourth, so other models still rank above it in this area. Google says it also worked on coding, agentic and scientific tasks, full-stack web development and multimodal understanding.

Carbon has the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. Cached input tokens are 95% cheaper. According to the Google blog post, the price rises to $4 per million input tokens and $20 per million output tokens after the introductory period.

Safety claims and the staged rollout

Google says Carbon uses the same layered safety stack as Argon. That stack combines model training, real-time misuse defenses, resilience against prompt injection and chain-of-thought monitoring. The company says it trained Carbon to respond correctly to "complex or high-pressure tasks". It also says Carbon showed lower misalignment and fabrication rates when its own teams used it in live coding environments.

The Google blog post gives three figures. Successful cheating on Terminal-Bench 4.0 fell to 0.0%. Unsupported claims fell by about half compared with Argon. Attack success on the Gray Swan indirect prompt injection benchmark was 0.0%. Argon's behaviour is documented in Google's Safety and Security report. The announcement does not link a separate model card for Carbon.

The order of the rollout follows from what the model is good at. A model that finds and exploits vulnerabilities is useful to defenders and attackers alike. The ExploitGym score measures exploitation, which is an offensive skill. Giving defenders access first, and keeping refining guardrails before a wider release, is how Google has chosen to handle that. All the numbers so far come from Google and its subsidiaries. No independent evaluation has been published yet.

What it means for work

If Google's description holds up, Carbon can take on whole security tasks that people usually do by hand: finding a vulnerability, confirming it can be exploited and writing the patch. The first people to test this are security engineers on defensive teams, followed by Fairwind partners. For software engineers, the change is in how long the model can work alone. A model that can spend days tracking down a subtle bug in a large or low-level codebase handles work that is normally given to senior engineers. The engineer's role shifts towards setting the task, reviewing the result and deciding whether a patch ships. Several things are still unknown. The internal results, such as the multi-day runs and the 10x Chrome comparison, have not been checked independently. Nobody yet knows how the model behaves on codebases outside Google. And there is no date for general access.