Google unveils Gemini 4 Carbon as a frontier model focused on cybersecurity, coding and long-running engineering work. The October 28 release starts with selected cyber defenders rather than a general product rollout, giving Google time to test safeguards before opening access to developers, businesses and consumers.

 

Carbon arrives two weeks after Gemini 4 Argon with higher company-reported scores across vulnerability discovery, exploitation and patching. Google also published introductory API pricing, but it has not announced a date for broad availability or identified every organization receiving early access.

 

The launch establishes four verified points:

  • Selected cyber defenders receive access first.
  • Trusted partners follow through Google's Fairwind Program.
  • Introductory pricing is $2 for input and $10 for output per million tokens.
  • Developer, business and consumer access remains forthcoming.

 

Further Reading

 

Google Unveils Gemini 4 Carbon With Restricted First Access

Google's launch announcement says Carbon is entering the U.S. government's voluntary pre-release testing process. A subset of cyber defenders will receive the model this week, followed gradually by other trusted partners participating in the Fairwind Program.

 

That sequence is central to the release. Carbon is not generally available in Gemini apps or through an open developer endpoint at launch. Google says it will collect feedback and refine guardrails before expanding the model across its product surfaces, without giving a firm date.

 

Reuters independently confirmed the staged rollout and introductory price. Input will cost $2 per million tokens and output $10 per million tokens during the introductory period. Cached input is priced 95% below the standard input rate.

 

Google says pricing will later rise to $4 per million input tokens and $20 per million output tokens. Those figures establish a planned commercial rate, but early cyber-defender access may operate under separate program terms that the announcement does not describe.

 

Cybersecurity Benchmarks Cover Detection, Exploitation and Patching

Carbon's most distinctive results span three stages of defensive security work. Google reports 88.2% on SEC-bench Pro for discovering vulnerabilities, 54.8% on ExploitGym for validating exploitability and 71.0% on CWE-bench for producing root-cause patches.

 

The DeepMind comparison table places those results above Gemini 4 Argon and the listed rival models on the same tests. Google explicitly says the cybersecurity evaluations ran without guardrails, meaning they measure underlying capability rather than the behavior users should expect under deployed access controls.

 

That distinction cuts both ways. A model that can locate and verify real flaws may help defenders prioritize repairs, but exploitation skill can also support offensive activity. Restricted access, identity checks, tool limits and monitoring therefore matter as much as the published benchmark score.

 

Google says Carbon more than doubled Argon's secure-code success rate on an internal benchmark. It also says Chrome engineers saw a more than tenfold patching improvement over an unnamed frontier model. Those results are useful directional evidence, but the missing test set and unnamed comparison limit independent interpretation.

 

Carbon Extends Long-Horizon Coding Beyond Security

Google positions Carbon as a broader engineering model rather than a single-purpose security scanner. Internal teams reportedly ran it for days on complex debugging and patching tasks. Such persistence could help with large repositories, provided the agent can recover from mistakes and stay within defined permissions.

 

On public comparisons, Carbon scored 61.2% on FrontierSWE 2 and 65.2% on Terminal-Bench 4.0 in Google's table. It also reached 59 on Artificial Analysis Intelligence Index v5.0, which Google describes as the highest result in that version.

 

Benchmarks compress a model's behavior into controlled tasks. They do not establish reliability across an organization's private code, build tools, credentials or deployment process. Early users will need to measure successful fixes, false positives, regressions and human-review time in their own environments.

 

Carbon's Safety Claims Need Deployment Evidence

Google reports that Carbon reduced unsupported claims by about half relative to Argon in internal testing. It also says successful shortcut use fell to zero on Terminal-Bench 4.0, while the model recorded a zero percent attack-success rate on the evaluated Gray Swan indirect prompt-injection set.

 

Zero on a test set does not mean prompt injection has been solved. Attackers can change instructions, data sources and tool chains after a benchmark is published. Security teams should treat the result as evidence about one evaluation and continue isolating agents from unnecessary networks, secrets and production permissions.

 

The initial rollout should generate more useful evidence than a broad launch would. Cyber defenders can test whether Carbon finds consequential flaws without flooding teams with low-value alerts, and whether its patches survive review and regression testing. Google can also observe misuse attempts before enabling wider access.

 

The next verifiable milestones are named Fairwind participants, wider API availability and independent reproduction of the security results. Until those arrive, Gemini 4 Carbon is a limited-access model with strong reported capabilities and defined prices, not a generally available replacement for existing coding or security tools.

 

Illustrative cybersecurity image by jaydeep_ via Wikimedia Commons, released under the CC0 1.0 public-domain dedication.