Gemini 4 Carbon: Google's new cyber-focused frontier model

Updated: November 10, 2026, 17:20 (Europe/Budapest): Google made Gemini 4 Carbon generally available to paying developers and enterprises. This article now includes the new independent benchmarks, the /swarm feature and updated safety data.

What happened

Google DeepMind announced Gemini 4 Carbon on October 28. Google calls it its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities (Google DeepMind).

Carbon builds on Gemini 4 Argon, which launched two weeks ago. According to Google, it ranks first on the recently updated v5.0 of the Artificial Analysis Intelligence Index with a score of 59. Google reports significant gains in coding, agentic tasks and capabilities critical for cyber defense.

Cybersecurity is the big jump

Google highlights three external benchmarks, all evaluated without guardrails:

• SEC-bench Pro, finding real vulnerabilities: 88.2%.
• ExploitGym, checking which vulnerabilities are exploitable: 54.8%.
• Collinear's CWE-bench, patching at the root cause: 71.0%.

On internal benchmarks, Carbon has more than a 2x higher success rate than Argon at generating secure code. The Chrome team reported that it performed more than 10x better than another lab's large frontier model at generating high-quality patches. Wiz also sees Carbon leading on web penetration testing benchmarks.

Source: Google DeepMind. Gemini 4 Carbon benchmark results alongside Gemini 4 Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1.

Even Google's own table is not all first places. GPT-6 Astra leads on FrontierSWE 2 (65.5% versus Carbon's 61.2%), and Claude Opus 5.5 leads on the OSWorld-2.1 offline subset (87.2% versus 81.5%).

Agent runs that last for days

Carbon is built for sustained agentic workflows. Google says that internally it ran for days, persisting through dead ends to root-cause and patch subtle bugs across large codebases and low-level systems.

Web development improved as well: Carbon gained 89 Elo points on WebDev Arena and sits in fourth place with 1769.

Safety and alignment

Carbon builds on Argon's multi-layered safety and security stack. Google reports these results:

• Terminal-Bench 4.0: successful cheating at 0.0% and cheating attempts at 0.4%, while overall accuracy improved.
• In live coding environments with internal teams, high-severity misalignment alerts dropped by almost 80%.
• In internal testing, it makes roughly half as many unsupported claims as Argon.
• Overreach fell by 26% and under-service by 29% on multi-step plans.
• Gray Swan IPI benchmark: 0.0% prompt injection attack success, down from Argon's 0.7%.

What changed on November 10

On November 10, Google opened Carbon to all paying developers and enterprises through the API in Google AI Studio and Gemini for business (Google DeepMind). It is also rolling out to Google AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search.

Google says Argon remains the right choice for most software engineering and knowledge work. Carbon is meant for the hardest tasks, such as complex multi-step coding, debugging and open-ended research. It uses more tokens than Argon to reason more thoroughly and run extended verification loops.

There are now independent results too. Google highlights these:

• CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench: first place, at more than 2x lower cost per task than the next best external model, according to Google.
• Artificial Analysis Multimodal Knowledge Agent (MKA), a new benchmark for dense documents with charts and drawings: second place, at more than 6x lower average cost per task than the leader.
• Balyasny Asset Management's financial benchmark: first out of 23 frontier model configurations.

Google Antigravity is previewing a /swarm command that turns a single prompt into a coordinated team of agents you can track in real time. Carbon takes the roles that need the deepest reasoning, such as architectural design, coordination and auditing, while faster, lighter models work in parallel isolated workspaces.

The cybersecurity results grew as well. In Project Fortify, Google's internal autonomous security system, Carbon hit 100% recall on scanning and 95% precision on triage on Google's internal benchmark. Wiz used it to find 0-day vulnerabilities at two major SaaS providers in under two hours.

New safety data arrived too. In simulated cyberattack environments, Carbon blocked 99% of attack trajectories from an adaptive red-teaming agent, and the per-attempt breach rate fell to 0.8% (3.1% on Argon). On a 375-task agentic honesty benchmark it passed 92% of tasks, compared with 70% for Argon.

Who gets access, and at what price

Google took a deliberate rollout and joined the US government's voluntary pre-release model testing process. For two weeks after October 28, only a subset of cyber defenders could use it through the Fairwind Program. Since November 10, every paying developer can access it through the API.

The introductory price matches Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper. After the introductory period, the price rises to $4 and $20. Google did not say how long the introductory period lasts.

What it means for you

You can try it now. The per-token price matches Argon during the introductory period, but Carbon uses more tokens per task, so the final bill may differ. Google itself recommends Argon for most work, so test Carbon on hard, multi-step coding and debugging tasks. The independent benchmarks look promising, but results on your own pipeline matter more.