Gemini 4 Carbon: what businesses should verify before access.
Google announced Gemini 4 Carbon for limited cyber-defender access. See what is known, what is missing and how businesses can prepare an evaluation.
Frontier AI · Current analysisBy Orbital Content Team · Published October 29, 2026
AI-assisted research and drafting with editorial review. Google and Google DeepMind sources were checked October 29, 2026. Reported model capabilities, benchmark results, internal observations, pricing and rollout plans remain Google statements unless explicitly labeled Orbital analysis. Orbital did not access Gemini 4 Carbon, apply to the Fairwind Program, call an API, inspect Google systems, reproduce benchmarks, validate security or safety claims, review customer-specific terms or provide offensive cybersecurity instructions.
Google announced Gemini 4 Carbon on October 28 as its strongest model for cybersecurity and long-horizon work. The company says access begins with a subset of cyber defenders and will expand gradually to additional trusted partners through its Fairwind Program before reaching developers, businesses and consumers.
That makes Carbon meaningful news, but not a generally available business product. Google's announcement does not publish a public model ID, API specification, region list, token limits, service level, product-by-product release date or general-access date. The responsible business response is therefore to prepare the decision and evidence requirements—not to plan a production migration around an endpoint that ordinary buyers cannot yet select.
What Google announced on October 28
The launch article describes Carbon as a frontier model built for sustained agentic work, including complex software engineering, full-stack development, multimodal understanding, scientific tasks and cybersecurity. Google says internal teams have seen the model persist for days while working through dead ends in large codebases and low-level systems. That is a vendor-reported observation, not a published service commitment or proof that a customer workflow should run unattended for days.
| Google statement | What it establishes | What it does not establish |
|---|---|---|
| Carbon was announced October 28, 2026. | The model and initial rollout plan are public. | Broad availability, a public endpoint or a production migration date. |
| Access begins with a subset of cyber defenders and expands through Fairwind. | Google is using staged, managed access for advanced cyber capability. | Eligibility, approval timing or access for a normal service business. |
| Google reports gains in coding, agents, science, multimodal work and cybersecurity. | The model is a candidate for future task-specific evaluation. | Results on another company's code, data, tools or acceptance criteria. |
| The announced launch price matches Argon's introductory price: $2 per million input tokens and $10 per million output tokens; cached input is described as 95% off. | Google has stated intended model-token rates. | That Carbon can be purchased today, the introductory end date, every billable unit or total workflow cost. |
| Google says later pricing will be $4 per million input and $20 per million output tokens. | A higher standard price should be included in future planning. | When a particular account will transition or which separate tools and services will be charged. |
Google's wording matters: Carbon will launch at the stated price and is coming to developers, businesses and consumers. A buyer should not copy the price into a production budget until the applicable product, contract, model ID, region, billing units and start date are documented for its account.
What a business still cannot verify publicly
As of October 29, Google's public model-card index lists Gemini 4 Argon but not Gemini 4 Carbon. The Carbon announcement and Gemini model page provide benchmark and rollout information, but the accessible primary sources do not establish a full API specification.
Before procurement or production use, request dated evidence for:
- public or allowlisted model ID and exact product surface;
- preview, general-availability or other lifecycle status;
- supported regions and data-processing location;
- input and output modalities, context window and output limits;
- tool calling, structured output, code execution, grounding and computer-use status;
- rate limits, quotas, concurrency, job duration, cancellation and retry behavior;
- service levels, support route, incident notices and maintenance;
- data use, retention, deletion, logging, access and training terms;
- abuse monitoring and eligibility for any reduced-retention control;
- security and safety report specific to Carbon;
- price effective dates, cached-input conditions, tool charges and regional differences;
- model update, alias, deprecation and migration policy; and
- customer responsibilities under Fairwind or another access program.
Absence from a public model-card index does not prove that no private documentation exists. It means Orbital could not verify a Carbon model card in the public index on October 29. A participant may receive additional terms or technical material that should be reviewed for the intended workload.
How to read Carbon's benchmark claims
Google reports that Carbon scored 59 on Artificial Analysis Intelligence Index v5.0, 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym and 71.0% on CWE-bench. The launch article explicitly says the three highlighted cyber evaluations were run without guardrails. That is important context: an unguarded research evaluation may measure capability, but it does not describe the behavior or access boundary of a customer-facing product.
The same article reports internal results: more than a twofold uplift over Argon on secure-code generation, more than a tenfold patch-quality difference against another lab's model in a Chrome example, nearly 80% fewer high-severity misalignment alerts, approximately half as many unsupported claims and zero successful attacks in the cited Gray Swan indirect-prompt-injection evaluation. These are Google-reported or Google-selected observations. Orbital did not reproduce them.
| Launch signal | Business test | Required evidence |
|---|---|---|
| Vulnerability discovery | Run authorized cases from the company's supported languages and repositories. | True findings, false positives, missed known cases, severity and reviewer time. |
| Exploit validation | Use isolated, authorized test environments and defensive scope only. | Permission, containment, reproducibility, audit record and no path to production targets. |
| Patch generation | Generate a standard diff, then run existing build, unit, security and regression tests. | Root cause addressed, tests pass, behavior preserved and human owner approves. |
| Long-horizon work | Give the model a bounded engineering task with checkpoints and a time budget. | Completion rate, elapsed time, tool calls, reversals, cost and intervention count. |
| Prompt-injection resistance | Expose the agent to approved malicious and conflicting content in a sandbox. | Denied actions, leaked data, tool misuse, escalation and recovery—not only a benchmark score. |
| Reduced unsupported claims | Require citations or repository evidence for factual and code assertions. | Supported-claim rate, missing-evidence behavior and reviewer corrections. |
Benchmark position can justify adding a model to an evaluation. It cannot replace evaluation on the company's codebase, data, tools, policies, latency and failure costs. A lower-scoring model may still be the better operating choice if it is available, cheaper, faster, easier to control or already validated for the task.
What Fairwind access means
Google's Carbon announcement says the model will expand through the Fairwind Program. Google's current Fairwind pages describe a limited-access program for governments and trusted partners, prioritizing governments, critical infrastructure operators and core technology platforms. Google says it has more than 650 partners globally.
The program page describes due diligence, managed access and restrictions to defensive and research purposes. It says participating organizations agree to controls such as user-level authentication, phishing-resistant multifactor authentication, applicable access controls, employee access tracking and limiting access to internal cybersecurity, incident-response or penetration-testing teams. Organizations may not share, redistribute or sell access.
Those pages still identify Gemini 4 Argon and earlier cyber models in several program details as of October 29. Carbon's launch article establishes the planned Fairwind rollout, but the public program page does not establish that every existing partner automatically receives Carbon or that the prior model's technical specifications carry over. An interested organization should contact its Google Cloud account team and request Carbon-specific eligibility, terms and documentation.
A normal service business that is not eligible has no reason to imitate sensitive cyber testing. It can prepare evaluation criteria and continue using publicly available models for low-risk, authorized tasks while it waits for Carbon's documented business release.
Choose a responsible first workload
Carbon's headline is cybersecurity, but a first business evaluation should still be bounded, reversible and reviewable. A suitable candidate has:
- a repository or synthetic test set the organization is authorized to use;
- a known expected result or reviewer rubric;
- no live customer, employee or credential data unless expressly approved for the documented service;
- read-only discovery or diff generation rather than automatic production changes;
- existing build and test automation;
- a named security or engineering reviewer;
- a manual or current-model baseline;
- an isolated execution environment and limited tools;
- a time, token and tool-call budget; and
- a stop condition and fallback.
Examples include reviewing intentionally vulnerable test code, explaining a known defect, prioritizing approved static-analysis findings or drafting a patch for human review in a non-production branch. Avoid first pilots that scan systems without authorization, operate against live targets, accept unfiltered secrets, merge code automatically, change accounts, send customer messages or make legal or compliance conclusions.
The AI automation readiness checklist can define the workflow, owner, data, tools, review and stop conditions before model access. The existing Gemini 4 Argon evaluation guide explains the broadly available predecessor and its documented API boundary.
Design the control boundary before connecting the model
A capable cyber model can help defenders and can create dual-use risk. Keep model access, code access, execution and deployment as separate permissions.
| Control | Decision to record | Acceptance evidence |
|---|---|---|
| Access | Named people, service identities, authentication and offboarding. | Least-privilege list, phishing-resistant MFA where required and revoked-user test. |
| Scope | Authorized repositories, systems, vulnerability classes and prohibited targets. | Written rules of engagement and denied out-of-scope test. |
| Data | Allowed, masked, prohibited and specialist-review information. | Synthetic or approved data set, input filter and exception route. |
| Tools | Read, execute, write, network and deployment permissions. | Allowlist, sandbox, scoped credentials and blocked overreach test. |
| Patch path | Branch creation, tests, review and merge authority. | Diff, test results, reviewer identity and separate human merge approval. |
| Logging | Requests, tool calls, outputs, approvals and data excluded from logs. | Trace from task to result with controlled access and retention. |
| Incident response | Disable access, preserve evidence, notify owners and restore workflow. | Tested off switch and tabletop result. |
| Change control | Model alias, version, prompt, tools and evaluation owner. | Version record and regression gate before changes reach production. |
Google says Carbon builds on Argon's layered defenses, including training, misuse defenses, prompt-injection resilience and chain-of-thought monitoring. That is relevant vendor information; it does not configure customer identity, repository permissions, network access, CI/CD approvals or incident response.
Model complete operating cost
Token price is one component. A security or engineering agent may read large repositories, reason for long periods, call tools, generate patches, retry after failures and require specialist review.
Total accepted-task cost = model input + cached input + thinking/output + tools and execution + storage and network + orchestration and logging + human review + failed attempts + remediation and incident overhead.
Collect these values during an evaluation:
- tokens and cache behavior for every run;
- model and tool calls per accepted result;
- elapsed time and queue time;
- compute used for builds, sandboxes and tests;
- false-positive investigation time;
- patch-review and correction time;
- failures, retries and abandoned runs;
- security, legal, procurement and platform work; and
- support, monitoring and recurring governance.
Calculate cost per accepted finding or approved patch, not per model call. Include the published standard rates in scenario planning. Do not assume the introductory rate will apply for a full year, and do not assume Argon's documented cache or tool behavior applies unchanged to Carbon until Carbon-specific documentation confirms it.
The AI automation ROI checklist provides a baseline, complete cost ledger, evidence ladder and stop-or-scale framework.
Build the evaluation scorecard now
An organization does not need Carbon access to define how it would judge Carbon. Create a fixed set of authorized, representative cases and score the current process, the current model and Carbon when access becomes available.
| Measure | Definition |
|---|---|
| Accepted-result rate | Cases meeting every required criterion after review. |
| Critical miss rate | Known important defects not found or not remediated correctly. |
| False-positive rate | Reported issues that do not meet the approved definition. |
| Patch validity | Builds, tests and security checks passed without unintended behavior change. |
| Unsupported claim rate | Assertions lacking repository, test or approved source evidence. |
| Boundary compliance | Out-of-scope data, tools or actions attempted. |
| Prompt-injection outcome | Unauthorized disclosure or action under the approved adversarial set. |
| Human effort | Review, correction, escalation and full-redo minutes. |
| Latency | Median and tail time for the complete workflow. |
| Total cost | All model, platform and labor cost per accepted result. |
Set thresholds before running Carbon. Score serious security failures separately from averages. Record the exact model identifier, lifecycle status, prompt, tools, configuration and date. A model update should trigger regression tests against the same set.
A bounded stop-or-expand pilot after access is documented
- Week 1 — verify the product. Review Carbon-specific eligibility, contract, lifecycle status, model ID, regions, data terms, price, quotas, support, logs, safety material and exit route. Do not substitute Argon documentation for missing Carbon evidence.
- Week 2 — run offline comparison. Use approved test repositories or synthetic cases. Keep tools read-only or limited to an isolated branch and compare against the current process.
- Week 3 — test controlled patches. Generate diffs, execute existing tests in a sandbox and require separate specialist review. Inject tool failures, conflicting evidence and approved prompt-injection cases.
- Week 4 — decide. Review accepted results, critical errors, boundary violations, latency, cost and reviewer effort. Stop, revise, retain as assistive or expand one controlled boundary.
Stop immediately for unauthorized target access, secret exposure, uncontained execution, repeated approval bypass, destructive changes, unexplained model/version change or unavailable audit evidence. Expansion should change one dimension at a time—repository, user group, data class, tool permission or traffic—and rerun the regression set.
Orbital's business take
Gemini 4 Carbon may become a significant option for demanding software-engineering and defensive-security work. The October 28 announcement gives businesses a reason to prepare an evaluation, not a reason to claim adoption.
The most important current facts are the staged access, the absence of broad product documentation and the distinction between capability benchmarks and an operating service. Security-sensitive teams should use the waiting period to define authorization, evidence, sandboxing, human approval, cost and off-switch requirements. Other teams can evaluate whether currently available Argon or a smaller model already satisfies the actual task.
Explore Orbital's AI automation services, tech stack consulting, connected-business planning, business consulting and service-business planning. A 60-minute consultation can turn one candidate workflow into an evaluation brief; model access, cloud usage, implementation, security, legal review and ongoing support are separately scoped.
Sources and editorial limits
Primary Google and Google DeepMind sources were checked October 29, 2026. Google announced Gemini 4 Carbon on October 28. Its launch article establishes the initial trusted-cyber-defender rollout, planned Fairwind expansion, vendor-reported benchmarks and intended token pricing. Google's Gemini model page and evaluation-methodology page provide current comparative results. The public model-card index did not list Carbon when checked. Fairwind pages describe the program's general eligibility, controls and existing model access; they do not establish Carbon access for every partner. Public pages, access, prices, terms and technical specifications can change. Orbital did not access or test Carbon, reproduce benchmarks, inspect Google infrastructure, review customer-specific terms or guarantee model quality, security, availability, compliance, savings or business results. This is business analysis, not cybersecurity, legal, privacy, procurement or financial advice.