Google has introduced Gemini 4 Carbon, a frontier AI model designed to improve software engineering, autonomous agent workflows and cybersecurity defense. The company describes Carbon as its strongest model to date, highlighting its ability to work on complex, multi-step engineering problems, discover software vulnerabilities and develop patches. Its release follows Gemini 4 Argon, which Google introduced approximately two weeks earlier.
The announcement reflects Google’s effort to build AI systems that can do more than answer questions or generate code snippets. Carbon is designed to sustain complex tasks, recover from unsuccessful approaches and work through large codebases over extended periods. However, access is initially restricted to selected cybersecurity defenders through Google’s Fairwind Program while the company continues testing its safety measures.
Key Takeaways
- New AI model: Gemini 4 Carbon was announced on October 28, 2026.
- Coding and reasoning: Google says the model is designed for complex, long-running engineering and agentic tasks.
- Cybersecurity performance: It scored 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, according to Google’s reported evaluations.
- Stronger safeguards: Google reports improvements in prompt-injection resistance, task alignment and reducing unsupported claims.
- Limited rollout: Initial access is restricted to selected cybersecurity defenders through the Fairwind Program.
- Pricing: Introductory API pricing is $2 per million input tokens and $10 per million output tokens, before standard rates of $4 and $20, respectively.
What Is Gemini 4 Carbon?
Gemini 4 Carbon is Google’s latest frontier model, built to handle demanding coding, scientific, multimodal and autonomous-agent workloads.
Unlike a conventional chatbot interaction, a long-running agentic task may involve examining a large codebase, identifying the cause of a bug, testing possible fixes and checking whether the final solution works. The model must maintain context, use tools effectively and adjust its approach when an initial attempt fails.
Google says Carbon can persist through difficult engineering tasks for days in internal testing. Examples include identifying subtle bugs and fixing problems across large codebases and low-level systems.
The company also reports improvements in full-stack web development and multimodal understanding. Carbon gained 89 Elo points in Google’s cited WebDev Arena evaluation, reaching fourth place with a score of 1,769.
These results suggest Google is targeting developers who need AI systems to complete larger portions of software work rather than simply produce isolated code suggestions. Actual results will depend on the task, development environment and tools available to the model.
Gemini 4 Carbon’s Cybersecurity Capabilities
Cybersecurity is the central differentiator in Google’s announcement. Carbon is designed to assist defenders across several stages of the security process: finding vulnerabilities, checking whether weaknesses can be exploited and developing patches.
Google reported the following benchmark results:
| Cybersecurity benchmark | What it evaluates | Gemini 4 Carbon |
|---|---|---|
| SEC-bench Pro | Finding vulnerabilities in real-world software | 88.2% |
| ExploitGym | Validating vulnerabilities through exploitation | 54.8% |
| CWE-bench | Developing fixes for software weaknesses | 71.0% |
The scores are based on Google’s reported evaluations. They should not be interpreted as guarantees that the model will achieve the same success rate in every production environment.
The distinction between identifying and exploiting a vulnerability is important. A potential weakness does not always translate into a practical attack. Security teams need to establish whether a vulnerability is real, understand its potential impact and prioritize remediation.
Carbon is designed to help with that process by connecting discovery, validation and patching. This could help teams investigate more findings and address weaknesses faster, although human review and conventional security testing remain important.
Google also says its internal benchmarks showed more than twice the success rate for generating secure code compared with Gemini 4 Argon. The company further reported that the Chrome team observed Carbon performing more than 10 times better than another frontier model from a different AI lab when generating high-quality patches. These are company-reported comparisons, not independent proof of universal superiority.
How Gemini 4 Carbon Compares With Gemini 4 Argon
Carbon builds on the capabilities of Gemini 4 Argon, Google’s previous frontier model, while emphasizing coding performance, agentic workflows and cybersecurity.
| Feature | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| Main emphasis | Advanced coding, long-running agents and cybersecurity | General frontier capabilities and agentic workflows |
| SEC-bench Pro | 88.2% | 83.4% |
| ExploitGym | 54.8% | 42.0% |
| CWE-bench | 71.0% | 68.0% |
| Terminal-Bench 4.0 | 65.2% | 57.1% |
| Initial cybersecurity access | Selected defenders through Fairwind | Earlier staged rollout |
Source: Google DeepMind’s Gemini model benchmarks.
These figures indicate improvements across several reported evaluations, particularly vulnerability validation and terminal-based coding tasks. However, benchmarks measure specific tasks under defined conditions. They cannot fully capture how a model will perform across every programming language, security environment or business workflow.
For developers, the practical question is whether Carbon can reduce the time required to solve difficult engineering problems while producing reliable, maintainable code.
Why Google Is Limiting Access to Cybersecurity Defenders
Powerful cybersecurity capabilities can have both defensive and offensive applications. A model that helps a security team find vulnerabilities in its own systems could also be misused to locate weaknesses in systems the user does not own or have permission to test.
Google is therefore introducing Carbon through its Fairwind Program, beginning with a subset of cybersecurity defenders and gradually expanding access to trusted partners. The company says it is also participating in the US government’s voluntary process for pre-release model testing.
The staged rollout is intended to give Google time to gather feedback and improve safeguards before broader availability to developers, businesses and consumers.
Google says Carbon uses a layered safety approach that includes model training, real-time misuse defenses, prompt-injection resistance and monitoring of model reasoning-related signals. The company reports that Carbon achieved a 0.0% attack success rate on the evaluated Gray Swan Indirect Prompt Injection benchmark, compared with 0.7% for Argon.
That result is encouraging, but it applies to the specific attacks and conditions tested. No benchmark can establish that a model is immune to every prompt injection or future attack technique.
Improvements in Reliability and Agent Behavior
Google’s announcement also emphasizes reducing errors that can undermine complex agent workflows.
The company says Carbon approximately halves unsupported claims compared with Argon in its internal testing. It also reports an almost 80% reduction in high-severity misalignment alerts in live coding environments used by its teams.
Google further reports a 26% reduction in task overreach and a 29% reduction in under-service. These terms refer to two different problems: going beyond the task a user authorized and failing to complete the requested work adequately.
Such improvements matter because a capable coding agent must do more than generate technically plausible output. It should also recognize ambiguity, identify missing evidence and respect the boundaries of the task.
The company says Carbon is more likely to surface conflicting information and explain what remains unresolved rather than fabricate results. These are vendor-reported findings, and broader independent evaluations will be useful in determining how consistently the improvements translate to real-world use.
Gemini 4 Carbon Pricing and Availability
Google plans to introduce Carbon at the same introductory API prices used for Argon.
| API usage | Introductory price | Standard price after the introductory period |
|---|---|---|
| Input tokens, per 1 million | $2 | $4 |
| Output tokens, per 1 million | $10 | $20 |
| Cached input tokens | 95% discount on input pricing | Subject to applicable pricing terms |
The introductory price is temporary. Google says the standard rates will apply after the introductory period, but its announcement does not specify a firm end date for Carbon’s introductory pricing.
At launch, Carbon is not immediately available to everyone. Initial access is being provided to selected cybersecurity defenders, with broader availability planned for developers, businesses and consumers through Google’s product offerings.
For companies evaluating the model, pricing is only one part of the calculation. The cost of running complex agent workflows also depends on token consumption, tool calls, task duration and the amount of human review required.
The Bigger Picture
Gemini 4 Carbon highlights the increasingly close relationship between advanced AI coding systems and cybersecurity. Models that can reason through large codebases and work autonomously for extended periods may help developers identify defects earlier, repair vulnerabilities and improve software quality.
Those same capabilities make controlled deployment essential. The challenge for Google is to give defenders access to powerful security tools while reducing the likelihood that malicious users can exploit them. The staged rollout reflects an attempt to balance these competing demands.
Competition among frontier AI developers is also shifting toward task completion, reliability and specialized performance. If Carbon’s reported gains hold up in independent testing and production deployments, it could strengthen Google’s position in coding and security-focused AI. Its long-term impact will depend on how well the model performs outside benchmark environments and how effectively its safeguards operate in practice.
FAQs
1. What is Gemini 4 Carbon?
Gemini 4 Carbon is Google’s frontier AI model announced on October 28, 2026. It focuses on advanced coding, long-running agentic tasks, scientific work and cybersecurity defense.
2. How good is Gemini 4 Carbon at cybersecurity?
Google reports scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. These are results from specific benchmark evaluations, not guarantees of performance in every real-world environment.
3. Can the public use Gemini 4 Carbon now?
Not yet through a general public release. Google is initially providing access to selected cybersecurity defenders through its Fairwind Program and plans to expand availability gradually.
4. How much will Gemini 4 Carbon cost?
Introductory API pricing is $2 per million input tokens and $10 per million output tokens. Google says standard pricing will later rise to $4 per million input tokens and $20 per million output tokens.
5. Is Gemini 4 Carbon safer than Gemini 4 Argon?
Google reports improvements in prompt-injection resistance, unsupported claims and task alignment. These results are based on its own testing, so independent evaluations and real-world experience will be important for assessing the model’s safety.
Looking Ahead
The next milestone for Gemini 4 Carbon will be its expansion beyond the initial group of cybersecurity defenders. Developers and businesses will want to evaluate whether its coding and security capabilities translate into faster development, better vulnerability remediation and fewer errors during extended agent workflows.
The broader question is whether frontier AI models can combine increasingly powerful technical capabilities with dependable safeguards. Carbon represents Google’s latest attempt to make progress on both fronts, but its ultimate value will be determined by independent evaluations, practical deployment results and the effectiveness of its security controls.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
