Home Gemini 4 Carbon
Limited access · Announced October 28, 2026
Gemini 4 Carbon: Benchmarks, Access and Pricing in One Place
Google DeepMind’s model for long-horizon engineering and cybersecurity defense is rolling out in stages. Track what is known about access, pricing and performance, with the original source linked.
Unofficial tracker · Last verified October 30, 2026

01 · Access
Can you use Gemini 4 Carbon yet?
General access has not been announced. Google said it would start with a subset of cyber defenders, then gradually expand to trusted partners through the Fairwind Program. Developer, business and consumer access is planned, but undated.
| Audience | Status in Google’s announcement |
|---|---|
| Cyber defenders | Initial access beginning the week of October 28 |
| Trusted partners | Gradual expansion through the Fairwind Program |
| Developers, businesses and consumers | Planned; no date announced |
Google also describes voluntary U.S. government pre-release testing; this is separate from public availability. Google announcement
02 · Model overview
What is Gemini 4 Carbon?
Carbon builds on Gemini 4 Argon, with a focus on sustained engineering and the cybersecurity defense lifecycle. Google says its internal teams have seen it run for days on complex, multi-step engineering tasks.
Long-horizon engineering
Multi-step debugging, root-cause analysis and code changes across large systems.
Cyber defense
Finding vulnerabilities, assessing exploitability and generating patches.
03 · Performance
Benchmarks reported by Google
These are the results Google selected for its announcement. Google says the cybersecurity benchmarks were evaluated without guardrails. Rankings and methodology differ between independent evaluations.
| Benchmark | Result | What it measures |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | 59 | Overall intelligence, ranked #1 in Google’s announcement |
| SEC-bench Pro | 88.2% | Finding real vulnerabilities |
| ExploitGym | 54.8% | Validating exploitability |
| CWE-bench | 71.0% | Patching vulnerabilities |
| WebDev Arena | 1,768 | #4 at announcement; 89 Elo above Argon |
Compare evaluation methods directly at Artificial Analysis and Vals AI. A rank on one index does not establish a rank on another. Artificial Analysis · Vals AI
04 · Pricing
What does Gemini 4 Carbon cost?
Google’s announced prices are per million tokens. The end date for the introductory period was not specified.
| Token type | Introductory | After introductory period |
|---|---|---|
| Input | $2 | $4 |
| Output | $10 | $20 |
| Cached input | 95% off input price | 95% off input price |
05 · Specifications
Model specifications
Artificial Analysis lists a 1 million token context window, text, image, speech and video input, text output, and proprietary weights. The maximum output length was not independently confirmed here.
Artificial Analysis model profile06 · Safety
Safety and alignment
Google reports fewer unsupported claims and fewer high-severity misalignment alerts in its internal testing than with Argon. It also reports a 0.0% attack success rate on the evaluated Gray Swan indirect prompt injection benchmark. These are test results, not a guarantee of safety.
07 · Questions
Frequently asked questions
Can I use Gemini 4 Carbon today?
Google announced that access would begin with a subset of cyber defenders during the week of October 28, 2026, followed by a gradual expansion to trusted Fairwind Program partners. The announcement did not provide a general release date for developers, businesses, or consumers. Availability may change, so check Google’s latest announcements before planning a project around access.
When was Gemini 4 Carbon announced?
Google announced Gemini 4 Carbon on October 28, 2026. Its initial announcement described a staged rollout, rather than immediate availability to everyone. The information on this page was last verified on October 30, 2026; later announcements may provide more detail about access, supported products, or pricing.
How much does Gemini 4 Carbon cost?
Google announced introductory prices of $2 per million input tokens and $10 per million output tokens. After the introductory period, the announced rates are $4 and $20 respectively. Cached input is priced at 95% off the input price. These are token rates, not a monthly subscription price, and your actual cost depends on usage and the applicable pricing period.
What is Gemini 4 Carbon designed for?
Google describes Gemini 4 Carbon as a model for long-horizon engineering and cybersecurity defense. The announcement highlights work such as finding vulnerabilities, validating whether they are exploitable, and helping patch them. Those are examples of intended applications, not a guarantee that the model can independently secure every system or complete every engineering task.
Can I use Carbon in the Gemini app or AI Studio?
Google says it plans to bring Carbon to developers, businesses, and consumers through its product surfaces. However, the announcement does not name individual products or give a release date for the Gemini app or AI Studio. Do not assume that access to another Gemini model automatically includes access to Gemini 4 Carbon.
How is Gemini 4 Carbon different from Gemini 4 Argon?
Google presents Carbon with a particular emphasis on long-horizon engineering and cybersecurity defense. In the announcement, Carbon scored 1,768 on WebDev Arena, 89 Elo above Argon. That result describes performance on one evaluation, not every possible task. Compare the models using your own prompts, quality requirements, costs, and access options when both are available to you.
What benchmarks are available for Gemini 4 Carbon?
Google reported a score of 59 on the Artificial Analysis Intelligence Index v5.0, 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench. It also reported 1,768 on WebDev Arena. These benchmarks measure different tasks, so they should be read individually rather than combined into a single promise of real-world performance.
What do the cybersecurity benchmark scores mean?
The evaluations highlighted in Google’s announcement cover different parts of a security workflow. SEC-bench Pro concerns finding vulnerabilities, ExploitGym concerns validating exploitability, and CWE-bench concerns patching. A strong result on one does not guarantee the same result on another, and benchmark conditions may differ from your codebase, tools, permissions, and threat model.
Can Gemini 4 Carbon review code for security issues?
Finding and patching vulnerabilities are among the defensive uses Google highlights. A practical review can still require human judgment, testing, and knowledge of the application’s architecture. Treat any model-generated finding as a lead to investigate, and verify suggested fixes against your own tests and security requirements before deploying them.
Is Gemini 4 Carbon suitable for web development?
Google included a WebDev Arena result of 1,768 in its announcement, providing one point of reference for web development tasks. A benchmark cannot tell you whether the model will meet every project’s accessibility, visual design, maintainability, or security requirements. Evaluate generated work in the context of your own application and review changes before release.
What is the context window for Gemini 4 Carbon?
Artificial Analysis lists a one-million-token context window for Gemini 4 Carbon. A context window is the amount of information a model can consider within a request; it is not the same as a guaranteed output length. Check the documentation for the specific product or API you use, because practical limits and supported features can depend on the integration.
What types of input and output does Carbon support?
Artificial Analysis lists text, image, speech, and video as input modalities, with text output. That model-level listing does not necessarily mean every interface will expose every input type. Before building a workflow, confirm which modalities are supported by the product, endpoint, and access tier available to you.
Is Gemini 4 Carbon open source?
Artificial Analysis describes the model weights as proprietary. That is different from a model whose weights are publicly released for download and self-hosting. For current licensing terms, deployment options, and permitted uses, consult the official documentation for the Google product or service through which access is offered.
How should teams evaluate Gemini 4 Carbon?
Start with representative tasks and compare results against your existing workflow. For engineering work, examine correctness, test coverage, maintainability, and the time needed for review. For security work, check both missed issues and false positives. Include token usage and pricing in the evaluation, and avoid sharing sensitive material until you have reviewed the relevant service terms and data-handling policies.
Where can I find reliable updates about Gemini 4 Carbon?
Start with Google’s original announcement, linked on this page, and check its later product and developer announcements for changes. Independent model profiles can provide another reference for specifications and evaluations. Pay attention to publication dates: access, product availability, and prices may change after the information on this page was last verified.
08 · Sources and updates
Keep following the Gemini 4 series
Read the original announcement for the full methodology and rollout details. For broader tracker updates, visit the Argon homepage.
