Gemini 4 Carbon: Powerful AI Model Tops Key Benchmarks

Google also reports gains in coding, multimodal understanding, scientific reasoning and complex knowledge work. Its published comparisons show Carbon ahead of Argon on many measures, although results vary across individual benchmarks.

Independent testing and real-world use will help establish how these results translate into everyday software development and business workflows.

3. Cybersecurity Is Gemini 4 Carbon’s Biggest Focus

One of the most important parts of Google’s announcement is Carbon’s performance on cybersecurity evaluations. Rather than concentrating only on generating code, the model is designed to assist with identifying weaknesses, evaluating whether they can be exploited and helping developers repair them.

Google reports the following results:

These results cover different stages of the cybersecurity process. Finding a potential vulnerability is not the same as proving that it can be exploited, and demonstrating exploitability is not the same as fixing the underlying issue.

Carbon’s reported scores suggest that Google is targeting the full defensive workflow, from identifying flaws to helping developers address them.

Google also says Carbon achieved more than twice the success rate of Argon at generating secure code on its internal benchmarks. The company reports improvements in patch quality and performance on complex engineering tasks.

These are Google’s reported results, not a guarantee that Carbon will discover every vulnerability or produce a correct patch. Security teams still need to validate findings, test fixes and review changes before deploying them.

4. Longer Tasks: Moving Beyond Simple AI Responses

Many AI assistants perform well on isolated questions but can struggle when a task requires repeated planning, debugging and revision.

Google says Gemini 4 Carbon is designed for long-horizon tasks—work that can require multiple stages and sustained effort. Its internal teams have reportedly seen the model work for days on complex engineering problems, including root-causing subtle bugs and patching issues across large codebases.

This capability could be useful in several areas:

• Software development: Investigating bugs that span multiple files and components.
• Testing: Helping create, run and interpret tests across a large project.
• Code maintenance: Understanding older systems before proposing changes.
• Security engineering: Investigating vulnerabilities and developing candidate fixes.
• Technical research: Working through multi-step problems that need repeated evaluation.

The distinction is important: a long-running AI agent needs to do more than generate a plausible answer. It must track progress, handle failed attempts, use tools appropriately and avoid making unsupported assumptions.

Google says Carbon has improved on these behaviours compared with Argon. Independent evaluations will be needed to determine how consistently it delivers those gains outside Google’s own environment.

5. Google Says Carbon Has Stronger AI Safety Measures

More capable coding and cybersecurity models can also create additional risks if they are misused or behave unpredictably. Google says Carbon includes improvements to safety, reliability and resistance to prompt injection.

The company reports a 0.0% attack success rate on the evaluated Gray Swan Indirect Prompt Injection benchmark, compared with Argon’s previously reported 0.7%. It also says Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts to cheat falling to 0.4%.

Google further reports fewer unsupported claims and better handling of ambiguous instructions. According to the company, Carbon is more likely to retrieve relevant sources, identify conflicting evidence and explain what remains uncertain rather than inventing missing details.

These results are encouraging, but they should not be interpreted as proof that the model is immune to prompt injection or every form of misuse. A zero score on a particular evaluation means no successful attacks were recorded under that test’s conditions; it does not establish universal security.

For organisations considering advanced AI agents, the model’s behaviour matters alongside the tools, permissions and monitoring around it.

6. When Will Gemini 4 Carbon Be Available?

Gemini 4 Carbon is being released gradually. Google says access begins with a subset of trusted cybersecurity defenders through its Fairwind Program. The company plans to expand access to additional trusted partners before making the model available more broadly to developers, businesses and consumers.

That means the announcement is not the same as a general public launch.

Google has also said Carbon is being tested through a voluntary pre-release model-testing process involving the U.S. government. The company says it will continue evaluating safeguards and gathering feedback as access expands.

If you want to try Carbon for everyday writing, research or coding, it is worth waiting for an official availability announcement instead of assuming that it is already accessible through the regular Gemini app.

Google has announced introductory API pricing of $2 per million input tokens and $10 per million output tokens, matching the introductory pricing for Gemini 4 Argon. Cached input tokens receive a 95% discount compared with the input-token price.

Google says the price will rise after the introductory period to $4 per million input tokens and $20 per million output tokens.

These are token-based API prices, not a monthly subscription for the Gemini consumer app. The actual cost of a project depends on how much input and output the model processes, the use of cached inputs and any additional service charges.

Developers should check Google’s official pricing page before budgeting a production deployment, particularly because the model’s initial access is restricted.

8. Gemini 4 Carbon vs Gemini 4 Argon: What’s Different?

Gemini 4 Carbon builds on the Gemini 4 model family, but Google is positioning it for stronger performance in coding, cybersecurity and longer engineering workflows.

The company’s published benchmark table reports a score of 59 for Carbon and 50 for Argon on Artificial Analysis Intelligence Index v5.0. Carbon also scores higher than Argon on the cybersecurity benchmarks highlighted in Google’s announcement.

The numbers above come from Google’s published comparison. They should be read as benchmark results rather than proof that Carbon will outperform Argon on every real-world task.

The most practical difference is the intended focus: Carbon is being introduced as a stronger model for long-running technical work and cybersecurity defense, while access to it remains more restricted.

9. What Gemini 4 Carbon Means for the AI Industry

The release reflects a broader shift in AI development. Companies are competing not only to build models that answer questions accurately, but also systems that can use tools, carry out longer tasks and help solve complex technical problems.

Carbon’s cybersecurity focus is especially significant. AI models can help defenders identify weaknesses and write patches faster, but similar capabilities also make responsible deployment and access controls important.

Google says it is beginning with trusted cybersecurity professionals so that the model can be evaluated in controlled settings before wider availability. This phased approach could help the company identify weaknesses and refine safeguards as more users gain access.

For developers, businesses and ordinary users, the immediate takeaway is that Carbon’s public announcement is an early stage in its rollout—not a promise that everyone can use it today.

Conclusion

Gemini 4 Carbon is Google’s new frontier AI model focused on advanced coding, long-running engineering tasks and cybersecurity defense. Its reported score of 59 on Artificial Analysis Intelligence Index v5.0 and its results across vulnerability detection, exploitation and patching make it a significant new release.

However, two points matter for readers: the benchmark results are based on published evaluations, and access remains restricted to a limited group of trusted cybersecurity defenders. Google has not announced a specific date for general public availability.

For developers and security professionals, Carbon is worth watching. For everyday users, the next important milestone will be when Google opens broader access and independent testing shows how well the model performs in real-world tasks.

Frequently Asked Questions

1. What is Gemini 4 Carbon?

Gemini 4 Carbon is Google’s frontier AI model designed for complex coding, long-running engineering tasks and cybersecurity defense.

2. Is Gemini 4 Carbon available to everyone?

No. Google is initially rolling it out to a subset of trusted cybersecurity defenders through the Fairwind Program. Broader access is planned, but no specific public release date has been announced.

3. What is Gemini 4 Carbon’s benchmark score?

Google reports a score of 59 on Artificial Analysis Intelligence Index v5.0, placing Carbon first in the comparison published by Google.

4. Is Gemini 4 Carbon better than Gemini 4 Argon?

Google’s published results show Carbon ahead of Argon on several benchmarks, including the Intelligence Index and the three cybersecurity evaluations highlighted in its announcement. Performance can still vary by task.

5. What is Gemini 4 Carbon’s API price?

The introductory price is $2 per million input tokens and $10 per million output tokens. Google says the planned prices after the introductory period are $4 and $20, respectively.

6. Can Gemini 4 Carbon find and fix software vulnerabilities?

Google reports strong performance in vulnerability discovery, exploitability evaluation and patching. These capabilities can assist security teams, but findings and generated patches still require validation and human review.

7. Is Gemini 4 Carbon available in the regular Gemini app?

Google has not announced general consumer availability. The initial rollout is limited to trusted cybersecurity defenders, with wider access planned for a later date.

Suggestions :
GPT-6 Sol and Luna Launch With 50% Lower API Prices
Claude Opus 5.5 Launches With Lower Costs and Fable-Level Performance
Grok 4.7 Review: Pricing, Benchmarks, Features and How It Compares

Source :
Google X Account
Google DeepMind — Gemini 4 Carbon: Our Next Frontier Model
Google DeepMind — Gemini Models and Benchmark Results
Google Blog — Q3 2026 CEO Remarks