Gemini 4 Carbon Explained: Access, Benchmarks & Pricing

Google released Gemini 4 Carbon on 28 October 2026, just two weeks after Gemini 4 Argon. The model leads the Artificial Analysis Intelligence Index v5.0 with a score of 59 and posts strong results on cybersecurity benchmarks. Most people cannot use it yet. Access starts with a subset of cyber defenders through the Fairwind Program.

Carbon improves on Argon in coding, agentic tasks, and long-horizon work. Teams at Google report it can run for days on multi-step engineering problems such as finding and patching subtle bugs in large codebases. The model also shows better resistance to shortcuts and prompt injection.

What Gemini 4 Carbon Does Differently

Carbon builds directly on Argon. Google focused improvements on coding quality, web development artifacts, multimodal understanding, and scientific tasks. On WebDev Arena it gained 89 Elo points and landed at fourth place with a score of 1768.

Internal teams observed reduced misalignment and fabrication rates in live coding environments. Carbon cuts unsupported claims by roughly half compared with Argon. When information is missing it surfaces the gap instead of inventing an answer.

The model also handles high-pressure tasks with tighter alignment. Successful cheating on Terminal-Bench 4.0 dropped to 0.0 percent, and high-severity misalignment alerts fell by nearly 80 percent in internal use.

Cybersecurity Benchmark Results

Google published three external cybersecurity numbers, all measured without guardrails:

• SEC-bench Pro (finding real vulnerabilities): 88.2 percent
• ExploitGym (validating through exploitation): 54.8 percent
• CWE-bench (patching at the root cause): 71.0 percent

These figures place Carbon ahead of other frontier models on the tests Google highlighted. Internal benchmarks showed more than a 2x lift in success rate for generating secure code over Argon. The Chrome team reported that Carbon produced high-quality patches more than 10 times better than another lab’s large frontier model in their evaluation.

On Gray Swan’s Indirect Prompt Injection Benchmark, Carbon scored 0.0 percent attack success rate. Argon previously scored 0.7 percent.

Pricing and Token Costs

Carbon launches at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95 percent discount off the input rate. After the introductory period ends, the price moves to $4 per million input tokens and $20 per million output tokens.

The pricing matches Argon’s introductory rate. Developers who already use Argon on the Gemini API or Google AI Studio will see familiar unit costs once Carbon becomes available on those surfaces. Google has not published a model ID or public API endpoint for Carbon yet.

Who Gets Access First

Access begins this week with a subset of cyber defenders. Google will then phase additional trusted partners into the Fairwind Program. Broader availability for developers, businesses, and consumers comes after Google gathers feedback and iterates on guardrails. No public timeline has been announced.

The Fairwind Program prioritizes governments and national cyber authorities, critical infrastructure operators (healthcare, telecommunications, energy, financial networks), and core technology platforms. Academic labs focused on defensive benchmarking can also apply. Google conducts background checks on applicant organizations and reviews their security history and ethical operations.

Partners must use phishing-resistant multi-factor authentication. They can grant access only to internal cybersecurity, incident response, or penetration-testing teams. Sharing, redistributing, or selling access is not allowed. The program already works with more than 650 partners worldwide.

Organizations that meet the criteria contact their Google Cloud account team or use the interest form on the Fairwind page. Individuals and most companies are not eligible at this stage.

How Developers Can Prepare Now

Carbon is not in the Gemini API or Google AI Studio today. The practical steps are:

1. Keep an active paid Gemini API project or Google AI Studio account so you can switch models the day a Carbon endpoint appears.
2. Document your current Argon prompts and evaluation sets. Carbon’s improved honesty and reduced shortcut behavior may change how you structure multi-step agent workflows.
3. Review long-horizon task patterns. Internal reports show Carbon persists through dead ends for days on complex engineering work. Test similar multi-day coding or remediation flows on Argon so the hand-off is measurable.
4. Monitor the official Gemini models blog and the Fairwind Program page for model ID announcements and expanded eligibility notes.

If you already integrate Argon into Android apps or Firebase AI Logic, the same client libraries will support Carbon once it is listed. Check our earlier guide on adding Gemini 4 Argon to Android apps for the integration pattern that will carry forward.

Practical Tips for the Waiting Period

• Track Artificial Analysis and independent coding benchmarks weekly. Carbon’s 59 on the v5.0 Index is the current high score; later models may move the number.
• Build evaluation harnesses that measure both task success and misalignment signals. Carbon’s lower fabrication and shortcut rates are useful only if your tests catch those behaviors.
• Separate defensive cybersecurity use cases from general coding. Fairwind access is restricted to vetted defensive work. General development will arrive later through standard product surfaces.
• Watch for Vertex AI and Model Garden listings. OpenAI models recently entered Google Cloud Model Garden; Carbon may follow a similar enterprise path once guardrails stabilize.

Conclusion

Gemini 4 Carbon is Google’s strongest model as of late October 2026. It leads the Artificial Analysis Intelligence Index, improves cybersecurity defense scores, and handles multi-day agentic coding with fewer shortcuts. Access remains limited to a subset of Fairwind Program cyber defenders while Google finishes guardrail testing. Developers can prepare by maintaining API access, documenting Argon workflows, and watching official channels for the public model ID.

The introductory price matches Argon at $2/$10 per million tokens. Once Carbon reaches developers, the same client code used for Argon will apply. Until then, the model is available only to approved defensive partners.

Sources

• Google Blog: Introducing Gemini 4 Carbon (28 October 2026)
• Google DeepMind Fairwind Program page
• Artificial Analysis Intelligence Index v5.0
• SEC-bench Pro, ExploitGym, and CWE-bench results as published by Google