Cyber defense gets Google's Gemini 4 Carbon

Google on Wednesday announced Gemini 4 Carbon, its newest frontier model, built for cyber defense and long-running engineering work. The company said the model ranks first on the updated Artificial Analysis Intelligence Index, with a score of 59, ahead of Anthropic's Claude Opus 5.5 at 58 and OpenAI's GPT-6 Astra at 57, according to Google and the independent evaluator.

The model is designed to work through the full security cycle: finding vulnerabilities in software, checking whether they can actually be exploited, and producing fixes. Google said its internal teams watched the model run for days on single engineering tasks, pushing through dead ends to trace subtle bugs across large codebases and low-level systems. That persistence is the point of the release, according to the company's announcement: earlier models tended to stall when a task stretched across many steps, while this one is built to keep going.

The launch comes a fortnight after Google introduced Gemini 4 Argon, the previous model in the same family. The quick succession shows how fast the leading labs are now shipping frontier systems, with coding and security work as the main battleground. OpenAI and Anthropic have both made similar claims for their latest models in recent weeks, and the benchmark tables have become the public scoreboard for that contest.

Security test results

On external benchmarks, the model scored 88.2 percent on SEC-bench Pro, which tests finding real vulnerabilities in existing software, and 54.8 percent on ExploitGym, which tests whether a model can validate a flaw by exploiting it. On CWE-bench, which measures patching at the root cause rather than papering over a symptom, it reached 71.0 percent. Google said the external benchmarks were run without guardrails in place.

Internal testing showed further gains. Google reported that the model generates secure code at more than double the success rate of Argon, and that its Chrome team found it produced high-quality patches more than ten times better than another lab's large frontier model. Wiz, Google's cloud security platform, also found the model leading on benchmarks for web penetration testing, a core part of its Scan for Good initiative, according to the announcement.

Outside security, the model gained 89 Elo points on WebDev Arena to reach 1768, placing fourth on that leaderboard for web development. It does not lead everywhere: a rival OpenAI system stays ahead on FrontierSWE 2 and a rival Anthropic system on OSWorld-2.1, according to Dealroom's comparison of the published tables. Those gaps matter because buyers compare models on the specific work they need done, not on a single headline ranking.

A limited rollout, with safety checks first

Access to the model starts narrow. Google said it will make the model available this week to a subset of cyber defenders, then gradually extend access to additional trusted partners through its Fairwind Program. Developers, businesses, and consumers will get it later, with no date given. The company said it is taking part in the United States government's voluntary process for pre-release model testing while it gathers feedback and refines safeguards.

The caution reflects what the model can do in the wrong hands. A system skilled at finding and exploiting flaws could help attackers as well as defenders, and Google's own safety report for the model family describes a multi-layered stack covering training, real-time misuse defenses, prompt-injection resilience, and monitoring of the model's reasoning. In internal testing, cheating attempts on Terminal-Bench 4.0 fell to 0.0 percent, and the model resisted every evaluated attack in Gray Swan's indirect prompt-injection benchmark. High-severity misalignment alerts in live coding work dropped by almost 80 percent, while overreach fell by 26 percent and under-service by 29 percent, according to Google.

Price and what it signals

The model will launch at the same introductory price as Argon: two dollars per million input tokens and 10 dollars per million output tokens, with cached input priced at 95 percent below the input rate. The pricing keeps Google competitive with Anthropic and OpenAI at a time when all three labs are charging developers for access to their most capable systems.

For security teams, the immediate question is whether the benchmark gains hold up outside Google's labs. The headline figures are company-reported, and independent testing will stay limited while access is restricted to vetted defenders. Reuters reported the launch on Thursday, noting that the model steps up competition with OpenAI and Anthropic as the race to build more capable coding and cybersecurity systems intensifies. Until wider access arrives, the scores are a claim to test, not a verdict, and defenders will judge the model on the flaws it finds in their own code.

Related coverage: Northstar Security's emergence with funding for AI agent cyber risks and Ultra Security's agentic AI platform launch.

Sources: Google's announcement of Gemini 4 Carbon and Reuters coverage of the launch.