Google has unveiled Gemini 4 Carbon, its newest frontier AI model, positioning it as the company’s strongest system for cybersecurity defense, coding, and long-running agentic work. Announced on October 28, 2026, Carbon is initially being released to a limited group of cyber defenders rather than broadly to developers or consumers.blog
The launch arrives just two weeks after Google introduced Gemini 4 Argon and intensifies competition with OpenAI and Anthropic in high-end coding and security-focused AI. Google says Carbon leads recent cybersecurity benchmarks for finding vulnerabilities, validating exploits, and producing code patches.reuters+1
What Google announced
Gemini 4 Carbon is designed for sustained, multi-step engineering work, including debugging large codebases, identifying security flaws, and generating fixes. Google describes it as its strongest model for “long horizon” tasks—work that may require an AI system to persist through setbacks over an extended period.blog
Koray Kavukcuoglu, senior vice president at Google DeepMind and Google’s chief AI architect, announced the model in Google’s official blog post. Google said Carbon is being introduced through a deliberate rollout, with access beginning for a subset of cyber defenders and expanding gradually to trusted partners through the company’s Fairwind Program.blog
Google has not yet given a public date for when Carbon will become available to developers, businesses, or consumers through Gemini product surfaces. The company said it intends to make the model available “as soon as possible” after gathering feedback and refining safeguards.blog
Cybersecurity benchmarks
The central claim behind Gemini 4 Carbon is its performance across the cybersecurity defense lifecycle: vulnerability discovery, exploit validation, and remediation. Google reported state-of-the-art results on three external benchmarks, evaluated without guardrails.blog
| Benchmark | Purpose | Gemini 4 Carbon result |
|---|---|---|
| SEC-bench Pro | Finding real software vulnerabilities | 88.2% |
| ExploitGym | Validating whether vulnerabilities are exploitable | 54.8% |
| CWE-bench | Producing fixes for identified weaknesses | 71.0% |
Google said Carbon achieved 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on Collinear’s CWE-bench. These results are company-reported benchmark scores, not independent audit findings, although Google says the external benchmarks were evaluated without guardrails.blog
What the scores mean
SEC-bench Pro: Measures a model’s ability to uncover complex vulnerabilities in real-world software.
ExploitGym: Tests whether a model can help determine which vulnerabilities are practically exploitable, helping defenders prioritize remediation.
CWE-bench: Evaluates whether a model can turn a validated finding into a working code fix that addresses the root cause.blog
Google also cited internal results: Carbon reportedly delivered more than twice the success rate of Argon in generating secure code, and Google’s Chrome team said Carbon performed more than 10 times better than another lab’s large frontier model when producing high-quality patches. Those comparisons are internal claims and should be treated as vendor-reported results rather than independently verified findings.blog
Coding and agentic performance
Beyond security, Google is presenting Carbon as a model for demanding software engineering. The company said its teams have used Carbon in live coding environments for tasks such as root-causing and patching subtle bugs across large codebases and low-level systems.blog
Carbon ranks fourth on WebDev Arena with an Elo score of 1,768, an increase of 89 Elo points, according to Google. The company also said Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59.blog
The model is intended to improve performance in coding, agentic workflows, scientific tasks, full-stack web development, and multimodal understanding. Google says these gains build on feedback from Gemini 4 Argon users.blog
Pricing and availability
Google said Gemini 4 Carbon will launch at the same introductory price as Gemini 4 Argon:
$2 per million input tokens
$10 per million output tokens
Cached input tokens priced 95% below the standard input-token price
After the introductory period ends, Google says pricing will rise to $4 per million input tokens and $20 per million output tokens.blog
Availability is currently restricted. Google said it is participating in the U.S. government’s voluntary pre-release model testing process and will begin by giving Carbon to a subset of cyber defenders this week. Broader access for developers, businesses, and consumers has not yet opened.blog
Safety and alignment claims
Google says Carbon uses the same layered safety approach as Argon, combining model training, real-time misuse defenses, prompt-injection resilience, and chain-of-thought monitoring. The company also says it trained Carbon to respond more reliably to complex or high-pressure tasks.blog
Among the safety-related claims, Google reported:
Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, while cheating attempts fell to 0.4%.
High-severity misalignment alerts in internal coding environments dropped by almost 80%.
Unsupported claims fell by approximately 2 times compared with Argon.
The model reduced overreach by 26% and under-service by 29%.
Carbon recorded a 0.0% attack success rate on Gray Swan’s Indirect Prompt Injection benchmark, compared with 0.7% for Argon.blog
These figures describe Google’s internal testing and benchmark results. They indicate the company’s safety goals and reported performance, but they do not eliminate the risks associated with deploying powerful AI systems in security-sensitive environments.
Why it matters
Gemini 4 Carbon reflects a broader shift in the AI industry toward models that can perform extended, tool-using work rather than simply answer prompts. For cybersecurity teams, the potential value lies in accelerating vulnerability research, triage, and patch development—areas where skilled professionals remain in short supply.blog
The launch also shows Google moving quickly within its own Gemini 4 family. Argon was introduced two weeks earlier, and Carbon now extends that line with a stronger emphasis on security and long-duration engineering tasks. Reuters characterized the release as part of an intensifying race among Google, OpenAI, and Anthropic to build more capable coding and cybersecurity systems.reuters
For organizations, the practical question is not only benchmark performance but also access, cost, integration, and safety controls. Because Carbon is not yet broadly available, most developers and enterprises will need to wait for Google’s wider rollout before evaluating it in production workflows.blog
What to watch next
Broader availability: Google has not announced a firm public release date for developers, businesses, or consumers.blog
Independent validation: Third-party testing will be important to confirm whether Carbon’s cybersecurity and coding results hold outside Google’s reported benchmarks.blog
Enterprise adoption: Access through the Fairwind Program may reveal how security teams use Carbon for vulnerability discovery, exploit validation, and patching.blog
Competitive response: OpenAI and Anthropic are likely to face pressure to match Google’s security-focused claims and pricing.reuters
Google’s announcement makes Gemini 4 Carbon one of the most security-focused frontier models released to date, but its real-world impact will depend on how well its benchmark gains translate into reliable, safe performance once access expands.blog
