Google announced Gemini 4 Carbon on October 28, 2026, with initial access limited to selected cybersecurity defenders rather than a general public release. The new model follows Argon and arrives with company claims of improvements in coding, autonomous tasks and cybersecurity. The Verge and Ars Technica separately reported the announcement and restricted rollout. (The Verge)
Google says access will begin with a subset of defenders this week, followed by additional trusted partners through its Fairwind Program. Availability for developers, businesses and consumers remains a plan, with no general-release date announced. (Google)
Access is the immediate constraint
For organizations outside that initial group, the announcement does not establish that Carbon is available to evaluate or deploy. Ars Technica reports that Google has made no commitment on the timing of a wide release. Expectations based on Argon's rollout should therefore not be treated as a Carbon launch schedule. (Ars Technica)
Fairwind predates Carbon. Google introduced the program on September 2 as a limited-access initiative for government agencies, trusted cloud customers and cybersecurity partners. Its stated purpose was to give defenders advanced capabilities to find and repair vulnerabilities, while adapting access and safeguards as the program developed. That background explains the distribution channel, not an entitlement to the new model. (Google)
The program's general website prioritizes governments, critical infrastructure operators and core technology platforms. However, the retrieved page still describes Argon in its model-specific access information. Those broader program descriptions should not be mistaken for a confirmed list of Carbon recipients or proof that every existing partner can use Carbon now. (Google DeepMind)
What Google says has improved
Google says its internal teams observed Carbon working for days on multi-step engineering problems, including diagnosing and patching bugs across large codebases. It also reports more than twice Argon's success rate on internal secure-code generation tests. These are company observations, not findings from hands-on testing by this publication. (Google)
For cybersecurity, Google reports 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. It explicitly says these evaluations were conducted without guardrails. The figures therefore should not be read as verified performance of the safeguarded service partners will receive. (Google)
That distinction separates capability from deployment: a test of what a model can accomplish without restrictions does not, by itself, establish how reliably it will perform under operational controls. The published security scores are evidence of Google's reported evaluation results, not an independent certification that Carbon can safely handle an organization's software unattended. (Google)
An independent lead, with a defined scope
There is independent support for the broader benchmark headline. Artificial Analysis lists Carbon at its High reasoning setting first on Intelligence Index v5.0 with 59 points, ahead of Claude Opus 5.5 at its Max setting with default fallback at 58. These are index points, not percentages or a measure of production reliability. (Artificial Analysis)
The evaluator's methodology combines nine assessments across four equally weighted categories: knowledge work, coding, science and general capability. It describes the suite as primarily text-based and English-language, with multilingual, speech and image capabilities evaluated separately. That makes the index a defined comparison of capabilities, rather than an all-purpose ranking of every task a business might need. (Artificial Analysis)
Artificial Analysis itself cautions that an aggregate score may not apply directly to every use case. For an enterprise choosing a model, the implication is straightforward: an overall lead is a reason to investigate relevant task results, not a substitute for evaluating the intended workflow. Nor does the index independently validate Google's separate cybersecurity claims. (Artificial Analysis)
Restricted deployment has industry precedents
Google is not alone in using selected partnerships for advanced cyber capabilities. In June, Anthropic announced an expansion of Project Glasswing from roughly 50 initial partners to approximately 150 additional organizations, with access contingent on meeting its security requirements. The program used Claude Mythos Preview to examine participating organizations' codebases for vulnerabilities. (Anthropic)
Anthropic said the expansion included sectors such as power, water, healthcare and communications. That provides a useful comparison: these programs concern software supporting essential services, not only developer productivity. It does not establish that the two companies' models have equivalent capabilities, safeguards or access conditions. (Anthropic)
For Carbon, the immediate practical distinction remains between a model announcement and a deployment option. Selected cybersecurity defenders are the initial audience; broader access has yet to receive a public timetable. Until that changes, most builders can examine the reported evidence but cannot treat Carbon as a generally available replacement for their current model. (The Verge)
