Gemini 4 Carbon: Benchmarks, Pricing & Access

Google announced Gemini 4 Carbon on October 28, 2026, positioning it as its strongest Gemini model for cybersecurity and long-horizon engineering work. The important buying detail is access: as of October 30, Google says Carbon is being rolled out first to a subset of cyber defenders and will expand to additional trusted partners before it reaches developers, businesses, and consumers more broadly. It is not a generally available model that every developer can call today.

This guide separates independently reported benchmark data from Google’s own claims, explains the announced token pricing, and outlines what security teams should verify before evaluating the model in a real environment.

Gemini 4 Carbon at a glance

What Gemini 4 Carbon is designed to do

Google describes Carbon as a model for sustained, multi-step work, including complex software engineering tasks that can involve investigating a difficult bug, tracing its root cause, and preparing a fix across a large codebase. The company also highlights cybersecurity defense work across vulnerability discovery, validation, and remediation.

Carbon builds on Gemini 4 Argon, but Google is positioning it as a more specialized step forward for high-stakes security and long-horizon tasks. That does not mean it should be treated as an autonomous security operator. Access, permissions, human review, and controlled testing remain important, especially because advanced vulnerability research can be dual-use.

Benchmarks: what the numbers actually show

There are two useful evidence layers. Artificial Analysis independently lists Gemini 4 Carbon (High) with a score of 59 and the top position in its Intelligence Index v5.0. Google separately reports the cybersecurity benchmark results below. These are different evaluations and should not be combined into one universal measure of model quality.

Google says its SEC-bench Pro, ExploitGym, and CWE-bench figures were evaluated without guardrails. That caveat matters: results from an unguarded benchmark are not a direct measure of how the model will behave in a production security workflow with policy controls, restricted tools, and human approval. The company also reports internal improvements in secure-code generation, prompt-injection resistance, and reliability. Those internal results should be treated as vendor-reported evidence until they can be independently reproduced.

Artificial Analysis lists the evaluated High variant with a 1-million-token context window and text, image, speech, and video inputs. Its current model page also shows no publicly available API. These details are useful reference points, but they are not a substitute for a published Google API specification or an account-specific access confirmation.

Availability: who can use it now?

Google’s October 28 announcement says Carbon is initially being made available to a subset of cyber defenders, with access gradually expanding to additional trusted partners through the Fairwind Program. Google says it intends to bring the model to developers, businesses, and consumers later, but it did not publish a general release date in that announcement.

The Fairwind Program is designed for high-priority defenders, such as governments, critical infrastructure operators, and core technology platforms. Google says it vets organizations and restricts use to defensive and research purposes. The program page explains its eligibility and governance approach; it should not be read as an open sign-up route for every individual developer.

If your organization is not eligible for the initial rollout, the practical next step is to monitor Google’s official model and developer announcements rather than relying on third-party access claims. Be cautious of unofficial endpoints that advertise immediate access to a model that Google has not made generally available.

Gemini 4 Carbon pricing

Google’s launch announcement listed introductory API pricing of $2 per 1 million input tokens and $10 per 1 million output tokens, with cached input tokens discounted by 95%. The same announcement states that pricing becomes $4 per 1 million input tokens and $20 per 1 million output tokens after the introductory period.

As checked on October 30, 2026, Artificial Analysis displays $4 input and $20 output per million tokens for Gemini 4 Carbon (High), while also marking the model as having no publicly available API. Because the model is still in a controlled rollout, treat these figures as announced or currently listed reference pricing—not a guarantee that a public account can purchase access today. Confirm the applicable rate, cache treatment, quotas, and billing terms when Google grants access.

For a practical budget, estimate input and output separately. Long-running coding and security workflows can generate substantial reasoning and output tokens, and retries or repeated repository context can increase spend. A useful starting point is Aiverio’s LLM API Cost Calculator; use the confirmed provider rates and realistic token counts rather than assuming the model’s benchmark score predicts cost per task.

Gemini 4 Carbon vs. Gemini 4 Argon

Google presents Carbon as building on Argon, with stronger emphasis on cybersecurity defense and sustained engineering tasks. Argon is the earlier Gemini 4 model already covered in Aiverio’s Gemini 4 Argon profile and Gemini 4 Argon vs. GPT-6.1 Sol comparison.

Carbon is not automatically the better choice for every workload. If your tasks are general writing, research, or routine coding, wait for public access and compare actual task quality, latency, and total cost against models you can already use. If your organization is eligible for the controlled rollout, prioritize security-specific evaluations and compare Carbon with Argon on the same approved test set.

What security teams should test before adoption

• Scope and authorization: restrict testing to systems and repositories your organization is authorized to assess.
• Human approval: require review before code changes, patches, or security actions are applied.
• Tool permissions: use least-privilege credentials, isolated environments, and explicit limits on network and repository access.
• Prompt-injection handling: test untrusted code, issue descriptions, and repository files as potential instruction sources.
• Patch quality: verify that proposed fixes address root causes and do not introduce regressions.
• Evidence and auditability: retain test cases, model outputs, reviewer decisions, and reproducible results.
• Cost and latency: measure tokens and wall-clock time per successful, reviewed task—not just per prompt.
• Data governance: confirm retention, data handling, access logging, and contractual protections before sending sensitive code or incident data.

Aiverio’s AI Browser Agent Security Checklist, AI Agent Governance Checklist, and Best LLM Observability Tools provide related controls for teams evaluating tool-using models and agent workflows.

Limitations and open questions

• General developer and consumer availability has not been announced with a specific date.
• Published benchmark scores do not guarantee accuracy or safe behavior on a company’s own systems.
• Several cybersecurity and alignment results in the launch post are Google’s own reported evaluations.
• Current public API availability and final billing terms remain uncertain while access is restricted.
• Security work requires authorization, human review, and controlled environments; benchmark performance does not replace a security process.

Frequently asked questions

Is Gemini 4 Carbon publicly available?

No general public release was announced on October 28, 2026. Google said it was starting with a subset of cyber defenders and would gradually expand access to trusted partners before broader availability.

What is Gemini 4 Carbon’s API price?

Google announced introductory pricing of $2 per million input tokens and $10 per million output tokens, followed by $4 and $20 after the introductory period. Artificial Analysis currently displays the $4/$20 rates and says no public API is available. Confirm the applicable rate when access is granted.

Is Gemini 4 Carbon better than Gemini 4 Argon?

Google positions Carbon as a stronger model for cybersecurity and long-horizon engineering, building on Argon. The better choice depends on the workload, access, quality on your own evaluations, latency, and total cost.

Can a regular developer apply for Carbon access?

The initial rollout is restricted. Google directs trusted partners through the Fairwind Program, which prioritizes defenders and critical organizations. A general developer sign-up date has not been announced.

Should companies use Carbon to run autonomous penetration tests?

Do not treat benchmark results as permission to run autonomous tests. Use only authorized environments, least-privilege access, human approval, and the program rules that apply to your organization.

Official and independent sources: Google’s October 28 Gemini 4 Carbon announcement, Artificial Analysis model evaluation, Artificial Analysis Intelligence Index methodology, and the Google DeepMind Fairwind Program.

Editorial note: Access, benchmark data, and pricing were checked against Google and Artificial Analysis sources on October 30, 2026. Benchmark claims are attributed to their publisher. Availability and prices can change; verify current access and billing terms before planning a deployment.