Google announced Gemini 4 Carbon on October 28. It plans access for selected cyber defenders that week, followed by trusted partners.

Google introduced Fairwind in September as a limited-access program for government, infrastructure and security partners. It described restrictions on who inside participating organizations could use the tools and requirements such as multifactor authentication. Those terms concern participation in the program; model-specific access can have its own timing and conditions.

Google described the program's first offering as a cyber model paired with its CodeMender harness for finding and fixing vulnerabilities. A model and the software that gives it tools are separate parts of that arrangement: a team considering access needs to know which capabilities, operational controls and deployment environment its own agreement covers. The September description is background to the program; access terms can differ by offering.

The big change

  • Google plans a phased Fairwind rollout, starting with a subset of cyber defenders. Most teams must wait before testing the model themselves.
  • Artificial Analysis's lead is tied to its measured configuration and mix of tasks; a buyer still needs evidence from its own workload.
  • Google reports an 88.2% result on the vulnerability-finding SEC-bench Pro without guardrails. Teams need to assess the safeguards and behavior of the version they can obtain.

What a score of 59 measures

Artificial Analysis lists Carbon at High reasoning with a score of 59. Claude Opus 5.5 at Max with its default fallback scores 58 on the same index. Its Intelligence Index v5 combines nine evaluations into four equally weighted categories: knowledge work, coding, science and general capability.

The index methodology describes a primarily English, text-based suite. Image, speech and multilingual evaluations are conducted separately. Within the coding category, Terminal-Bench 4.0 contributes 10% of the overall index and tests 66 terminal-based tasks. The composite may not predict a particular team's workload.

The Terminal-Bench maintainers' version notes explain why the version label matters. Version 4.0 changed resource limits, removed eight tasks and corrected others; its agent timeout is eight hours per task. Those changes affect the conditions under which a run produces its answer. A result from a different version is therefore not automatically a like-for-like comparison, even when both carry the Terminal-Bench name.

The maintainers say they changed instructions, environments and verifiers in 19 tasks to address problems found in earlier runs. That is a change to the test itself, not simply a new model taking the same exam. A team tracking a vendor's progress over time should keep the version with the score and check whether the underlying tasks changed before drawing a trend line.

Teams can test their own tasks against consistent criteria and record costs. BIG CHANGE has not tested Carbon.

Access and price remain distinct decisions

The SEC-bench Pro project builds cases from real security bugs with reproducible environments and validation artifacts. Its current repository spans browser engines and the Linux kernel; the benchmark can be updated over time. A benchmark's scope and settings shape what its score can say about real work, and results from different snapshots need careful comparison. Google has not published guarded scores in its announcement.

The repository describes case metadata, reproduction assets and validation logs, with fixed-image checks where applicable. These artifacts let an evaluator check a generated proof of concept against a known case rather than rely only on a model's explanation. That is useful evidence about the test, but a successful case does not show that every proposed fix is safe to deploy into an organization's production system.

An independent NIST assessment illustrates how the test setup affects interpretation. In its separate evaluation, the agent received source code with a known vulnerability and was pointed toward relevant files and the type of flaw. NIST also described turn limits and model settings. Those details apply to NIST's experiment; published scores need a comparable setup.

Google's announcement gives introductory prices of $2 per million input tokens and $10 per million output tokens, with cached input discounted 95%. It says prices become $4 and $20 after that period. It does not give an end date in the announcement.

Total task cost also depends on the tokens and attempts needed to complete it.