The Next News

// technology and science news, explained

AINews3 mins read6 sources

Google gives its new Gemini 4 Carbon model to cyber defenders before a wider release

Google says its strongest model leads benchmarks for finding, exploiting and patching software flaws. Vetted security teams get it first.

By AI Desk ·

Google on October 28, 2026, announced Gemini 4 Carbon, which it calls its strongest model and says has state-of-the-art cybersecurity and long-horizon capabilities. It is not yet releasing the model to the public. A select group of cyber defenders gets it first, from this week, and Google will widen access step by step to more trusted partners in its Fairwind Program. Developers, businesses and consumers will follow “as soon as possible,” the company says, once it has gathered more feedback and refined the model's guardrails.

Carbon follows Gemini 4 Argon, which Google announced on September 30 for cyber defenders and made generally available to developers, enterprises and Google AI Pro and Ultra subscribers on October 13. As with Argon, Google says it is taking part in the U.S. government's voluntary process for testing models before release.

Built to find and fix flaws

Google calls Carbon “a significant leap in cybersecurity capabilities.” It reports these results on external benchmarks, evaluated without guardrails:

Gemini 4 Carbon on cybersecurity benchmarks, as reported by Google (October 28, 2026)
BenchmarkTaskScore
SEC-bench ProFinding real vulnerabilities88.2%
ExploitGymExploiting vulnerabilities54.8%
CWE-bench (Collinear)Patching vulnerabilities at the root cause71.0%

In Google's internal tests, Carbon writes secure code successfully more than twice as often as Argon, the company says, and its Chrome team reported Carbon doing more than 10 times better at producing high-quality patches than a large frontier model from another, unnamed lab. Google set out the trade-off when it released Argon:

The same coding and scientific capabilities that help defenders patch hospital software or accelerate bioengineering also require strong safeguards against malicious actors.

Koray Kavukcuoglu, SVP, Google DeepMind, as Gemini 4 Argon became generally available on October 13, 2026

Who gets it first

Through Fairwind, defenders that Google considers high priority, such as governments, healthcare providers and telecommunications services, get its advanced models early, according to the program's page. The page still describes access to Argon and puts the number of partners worldwide at more than 650. Members must use user-level authentication and phishing-resistant multi-factor authentication, may give access only to internal cybersecurity, incident-response or penetration-testing teams, and may not share or resell it. Google says it runs background checks on applicants. With Argon, Fairwind partners received the model without its cyber guardrails.

Google did not name the U.S. agency involved in its pre-release testing. The National Institute of Standards and Technology's Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) says it is meant to be industry's main contact in the U.S. government for testing commercial systems. Its tasks include signing voluntary agreements with developers and running unclassified evaluations of capabilities that could threaten national security, with a focus on risks such as cybersecurity, biosecurity and chemical weapons.

Price and rankings

Carbon's launch price matches Argon's introductory rate: $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper. Once the introductory period ends, the price rises to $4 and $20.

Google says Carbon ranks first, with 59 points, on version 5.0 of the Artificial Analysis Intelligence Index, which combines nine evaluations. The firm's own leaderboard lists Carbon one point ahead of Anthropic's Claude Opus 5.5. Google also says Carbon climbed 89 Elo points to fourth place, at 1768, on the WebDev Arena leaderboard.

The company also reports better behavior on its own tests. Successful cheating on Terminal-Bench 4.0 fell to 0.0%, with attempts at 0.4%; high-severity misalignment alerts in live coding with Google teams fell by almost 80%; and unsupported claims fell by about half compared with Argon. Carbon resisted every attack in Gray Swan's indirect prompt injection benchmark, on which Argon's attack success rate was 0.7%. These are Google's own figures; its post points to the safety and security report it published for Argon.

What comes next

Google has not given a date for wider availability. With Argon, general availability came 13 days after the defenders-first announcement.

Sources

  1. Gemini 4 Carbon: our next frontier model Google blog.google
  2. Gemini 4 Argon: our next era of frontier intelligence Google blog.google
  3. Gemini 4 Argon: Built to solve your hardest problems Google blog.google
  4. Fairwind Program Google DeepMind deepmind.google
  5. Artificial Analysis Intelligence Index v5.0 Artificial Analysis artificialanalysis.ai
  6. Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) National Institute of Standards and Technology nist.gov

Checked against the 6 sources above on October 31, 2026. Spotted an error? Tell us.