Gemini 4 Carbon arrives for long-running agents and cybersecurity, defenders first

Google has introduced Gemini 4 Carbon, its new frontier model, in an October 28 announcement. According to the post by Koray Kavukcuoglu, SVP at Google DeepMind and chief AI architect, Carbon is the company's strongest model so far and is built for sustained agentic workflows and cybersecurity. It builds on Gemini 4 Argon, which Google unveiled on September 30 and opened to developers on October 13. For now, only cyber defenders get access.

Agents that keep going for days

Google says Carbon is designed for long, multi-step agentic work. Internal teams have seen the model run for days on complex engineering tasks, working its way past dead ends. On general benchmarks, Google says Carbon ranks first on the recently updated version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, and fourth on the WebDev Arena leaderboard for web development.

Three cybersecurity results

Cybersecurity is the centre of the announcement. In Google's own tests, run without guardrails, Carbon scored a state-of-the-art 88.2% on SEC-bench Pro, which measures vulnerability discovery. It set a new high of 54.8% on the exploitation benchmark ExploitGym and reached 71.0% on Collinear's CWE-bench, which measures patching. The announcement does not name the rival models behind these comparisons.

On internal benchmarks, Carbon more than doubled Argon's success rate at generating secure code. The Chrome team reported that it was more than 10 times better than another lab's large frontier model at producing high-quality patches; that model is not named. Cloud security firm Wiz says Carbon leads on web penetration testing benchmarks.

Safety and alignment figures

Google says Carbon uses the same layered safety stack as Argon: model training, real-time misuse defences, resilience to prompt injection and monitoring of the model's chain of thought. By the company's figures, successful cheating on Terminal-Bench 4.0 fell to 0.0%, with cheating attempts down to 0.4%. High-severity misalignment alerts in internal live coding use dropped by almost 80%. On Gray Swan's indirect prompt injection benchmark, attack success went from 0.7% for Argon to 0.0% for Carbon. Google also says unsupported claims fell by roughly half compared with Argon.

Defenders first

From this week, Carbon is available to a subset of cyber defenders. Google plans to widen access gradually to more trusted partners through the Fairwind Program, the controlled early access scheme it also used for Argon, which serves priority defenders such as governments and critical infrastructure operators. Google also says it is taking part in the US government's voluntary pre-release model testing. No date has been given for developers, businesses or consumers; Google says it will bring the model to them as soon as possible after gathering feedback and refining its guardrails. Current pricing is listed in Google's announcement.

For a fuller picture of the results, follow independent researchers' evaluations as well.

Sources: blog.google (Gemini 4 Carbon), blog.google (Gemini 4 Argon for developers)