Gemini 4 Carbon is Google's frontier model for cybersecurity and days-long agent runs
- Notes
- Cybersecurity
- AI agents
- Written by
- Bhargav Tibadiya
- Published
- Reading time
- 4 min read
Google announced Gemini 4 Carbon on October 28, 2026, and it is the strongest model Google has shipped. Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59, sets state-of-the-art results across the cybersecurity defense lifecycle, and is built to sustain agentic work over days rather than minutes. It arrives two weeks after Gemini 4 Argon, and Google is rolling it out deliberately: cyber defenders first, developers and consumers later.
What Gemini 4 Carbon is
Gemini 4 Carbon is Google DeepMind's next frontier model, announced by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google. The focus areas are cybersecurity defense and long-horizon agentic engineering: Google says its internal teams have watched Carbon run for days, pushing through dead ends to root-cause and patch subtle bugs across large codebases and low-level systems. On WebDev Arena, which ranks models on full-stack web development, Carbon gains 89 elo points over its predecessor to land at number 4 with a score of 1768.
How good is Gemini 4 Carbon at cybersecurity
Cybersecurity is the headline capability. Evaluated without guardrails, Carbon leads the benchmarks that map to each stage of the defense lifecycle:
- Finding vulnerabilities: 88.2 percent on SEC-bench Pro, a state-of-the-art score for autonomously uncovering complex flaws in real-world software.
- Validating exploitation: 54.8 percent on ExploitGym, the benchmark defenders use to check which vulnerabilities are actually exploitable.
- Patching root causes: 71.0 percent on Collinear's CWE-bench, turning validated findings into working code fixes.
The early field reports back the benchmarks up. On Google's internal benchmarks, Carbon more than doubles Argon's success rate at generating secure code. The Chrome team reports Carbon produced high-quality patches more than 10 times better than another lab's large frontier model, and Wiz, Google's cloud security platform, sees it leading on web penetration testing benchmarks.
How safe and aligned Gemini 4 Carbon is
Google paired the capability jump with measurable alignment gains, and the numbers are unusually specific:
- Successful cheating on Terminal-Bench 4.0 dropped to 0.0 percent, with cheat attempts down to 0.4 percent, while overall accuracy improved.
- High-severity misalignment alerts in live coding with internal teams fell by almost 80 percent.
- Unsupported claims in deliverables are cut by about half compared with Argon, and the model flags what is missing instead of fabricating around it.
- Overreach fell 26 percent and under-service 29 percent on multi-step plans.
- On Gray Swan's indirect prompt injection benchmark, Carbon resisted every evaluated attack, a 0.0 percent success rate where Argon scored 0.7.
Who can use Gemini 4 Carbon, and what it costs
Access is deliberately staged. Google is participating in the US government's voluntary pre-release testing process, and starting this week Carbon is available to a subset of cyber defenders. Access then phases to additional trusted partners through Google's Fairwind Program, with developers, businesses and consumers getting it through Google's product surfaces later, after more feedback and guardrail iteration.
Pricing matches Gemini 4 Argon's introductory rate: $2 per million input tokens and $10 per million output tokens, with cached input at 95 percent off. After the introductory period, the price rises to $4 per million input and $20 per million output tokens.
What Gemini 4 Carbon means for teams building software
Two things stand out if you build products for a living. First, the cybersecurity lifecycle numbers make defensive security work a realistic product feature rather than a specialist service: vulnerability discovery, exploit validation and patch generation each cleared a state-of-the-art bar in one model. Second, the long-horizon behavior is the part to watch. A model that sustains multi-day agent runs, cheats measurably less and resists prompt injection at 0.0 percent is a different foundation for unattended automation than anything priced at $2 per million input tokens before now. The obvious caveat is access: unless you are in the defender cohort or the Fairwind Program, this is a model to plan around, not to build on this week.
The cost side of that planning mirrors what I wrote about OpenAI's pricing move in the GPT-6.1 Terra note. The routing pattern behind features like these is covered in my Agent Skills case study.
FAQ
When was Gemini 4 Carbon released?
Google announced Gemini 4 Carbon on October 28, 2026, two weeks after Gemini 4 Argon. It is rolling out in stages: a subset of cyber defenders first, then trusted partners through the Fairwind Program, then developers, businesses and consumers.
What benchmarks does Gemini 4 Carbon lead?
Carbon leads the three benchmarks that map to the cyber defense lifecycle: 88.2 percent on SEC-bench Pro for finding vulnerabilities, 54.8 percent on ExploitGym for validating exploitation, and 71.0 percent on CWE-bench for patching root causes. It also ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59.
How much does Gemini 4 Carbon cost?
The introductory price matches Argon at $2 per million input tokens and $10 per million output tokens, with cached input 95 percent off. After the introductory period it rises to $4 per million input and $20 per million output tokens.
Is Gemini 4 Carbon available in the API yet?
Not yet for general developers. During the deliberate rollout, access goes to cyber defenders and Fairwind Program partners first. Google says it will bring Carbon to developers, businesses and consumers through its product surfaces as soon as possible.
Sources
- Announcement, benchmarks, safety results and pricing: Introducing Gemini 4 Carbon - Google (opens in a new tab)
Elsewhere on this site
The selected work catalogue lists the SaaS products and dashboards behind notes like this one.
