Google Gemini 4 Carbon Announced: Cybersecurity Benchmarks, Restricted Access and What Enterprises Should Know

28 October 2026 | Verified AI release analysis. Google has announced Gemini 4 Carbon, its latest frontier model, just two weeks after the broader release of Gemini 4 Argon. The headline numbers are striking, particularly for vulnerability discovery and code remediation. But there is an equally important detail for anyone planning an enterprise rollout: Gemini 4 Carbon is not yet generally available to developers, businesses or consumers. Google says a limited group of cybersecurity defenders will begin receiving access this week, with wider access to follow.

Google's original 28 October announcement describes Carbon as a model designed for sustained coding, cybersecurity and long-horizon agentic work. This is a genuine new model announcement, not an update to the already available Gemini 4 Argon.

What is actually new in Gemini 4 Carbon?

Google reports a score of 59 on version 5.0 of the Artificial Analysis Intelligence Index and says Carbon ranks first on that version of the index. The company also reports 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploit validation, and 71.0% on CWE-bench for software patching. These are reported evaluation results, not proof that an enterprise deployment will achieve those percentages or that every security finding is correct.

The benchmark names matter. Finding a flaw, demonstrating that it is exploitable and producing a safe patch are different tasks. A model that performs well on one may still require substantial human review on the others. Google says the external cyber benchmarks were evaluated without its normal guardrails, so results should not be treated as a direct measure of the experience available through a future public product.

Carbon is also aimed at multi-day engineering tasks. Google describes internal use cases in which the model continues through failed approaches to investigate and repair difficult bugs in large codebases. That may be valuable for complex modernization programs, but Google has not published a generally accessible deployment configuration that an independent enterprise team can reproduce today.

Is Gemini 4 Carbon available now?

Only to selected participants. Google says it will start making Carbon available to a subset of cybersecurity defenders this week and gradually expand access to trusted partners through its Google DeepMind Fairwind Program. Wider availability through developer, business and consumer products is planned, with no firm general-release date in the announcement. Existing Gemini subscriptions, an active Gemini API key or access to Gemini 4 Argon should not be assumed to include Carbon.

Google states an introductory API price of $2 per million input tokens and $10 per million output tokens, with cached input priced 95% lower. Its announcement says the price after the introductory period will be $4 per million input and $20 per million output tokens. This is announced future pricing, not evidence that ordinary API customers can call Carbon today. Verify the model identifier, billing, access and effective price when a supported product surface actually opens.

What do the security and alignment claims mean?

Google says its internal tests show more than twice Argon's success rate at producing secure code, a reduction in unsupported claims, fewer overreaching actions and stronger resistance to prompt injection. It also reports a 0.0% attack success rate on the evaluated Gray Swan indirect prompt-injection benchmark. A zero in a particular test set is not a guarantee of immunity to prompt injection in production. The model may encounter different tools, documents, adversaries, permissions and data flows in real enterprise applications.

For security leaders, this distinction is fundamental. Models can assist with finding and explaining flaws, but security teams must still validate severity, reproduce findings safely, review patches and test for regressions. A benchmark percentage does not replace threat modeling, independent penetration testing, code review or production change control.

How should enterprise architects evaluate Carbon when access opens?

1. Define a narrow task. Start with an internal codebase or controlled test repository and a measurable problem, such as reducing time to triage a known class of bugs. Avoid broad autonomous production permissions.
2. Set a comparable baseline. Run the same tasks against the current approved model and standard developer tools. Compare correctness, false positives, time, token usage and human corrections.
3. Constrain tools and data. Use least-privilege access, isolated execution, approved repositories, explicit network restrictions and reviewed pull requests. Keep secrets and regulated customer data out of prompts unless approved controls are in place.
4. Test failure paths. Include ambiguous tickets, malicious repository instructions, stale documentation, conflicting requirements and tasks where the correct response is to stop and request human authorization.
5. Measure full workflow cost. Account for human review, repeated attempts, testing infrastructure, input/output tokens and rework. A lower nominal token price does not necessarily make a long-running agent cheaper.

Why this matters for SAP and enterprise transformation teams

There is a credible future use case for AI-assisted analysis of custom developments, integration code and test automation. But Carbon's announcement does not mean it is embedded in SAP Joule, SAP BTP, SAP Build or any SAP SuccessFactors feature. No such integration or entitlement is established by Google's announcement. SAP customers should evaluate model access and approved data flows separately from any claims about SAP product availability.

For an SAP program, a useful experiment would compare proposed code corrections against released APIs, clean-core principles, unit tests and regression outcomes. A model-generated patch should enter the same transport, authorization and change-management process as any human-authored patch. The value is potentially faster analysis, not bypassing engineering governance.

Carbon versus Argon: do not confuse the releases

Gemini 4 Argon was introduced earlier and subsequently made available to developers and paying subscribers. Carbon is a separate, newer model with a restricted initial cybersecurity rollout. Google's Argon announcement provides the prior-generation context, but Carbon's product availability should be judged from the 28 October Carbon announcement, not from an Argon access page.

Bottom line: a model launch, not a universal product release

Gemini 4 Carbon deserves attention because Google is reporting substantial improvements in cybersecurity and sustained agentic coding. It does not yet justify claims that every enterprise can deploy it or that the reported benchmark results will translate directly into operational savings. The immediate action is to understand eligibility for restricted access and design an evaluation with defensible baselines, explicit security controls and measurable outcomes. When broader access is confirmed, teams can test whether the model delivers value in their own environment.

Primary sources and verification status

Source: Google's original Gemini 4 Carbon announcement, 28 October 2026. Context: Gemini 4 Argon announcement; Artificial Analysis benchmark index. All performance figures are attributed to Google's publication. Independent enterprise performance, universal availability and a firm general-release date remain unverified as of 28 October 2026.