Gemini 4 Carbon: Google's most powerful model launches for cyber defenders first

What this is about

On October 28, 2026, Google DeepMind announced Gemini 4 Carbon. The company calls it its most powerful model yet, with a special focus on cybersecurity and "long-horizon" tasks, meaning work that stretches over a very long time and many steps. Carbon builds on Gemini 4 Argon, which launched just two weeks earlier. The model isn't generally available yet. At first, only a select group of cyber defenders gets it. All figures in this article come from Google itself unless noted otherwise.

A phased rollout

Google describes the approach as "deliberate." The company takes part in the US government's voluntary process for pre-release model testing. This week, a subset of cyber defenders gets access, followed gradually by more trusted partners through the so-called Fairwind Program. In parallel, Google is gathering feedback and improving its safeguards ("guardrails") before developers, businesses and end users get access through Google's products. Google gives no specific date, only "as soon as possible."

Pricing is set to match Argon's launch pricing:

Tokens are the small chunks of text AI models split language into, and they're what you're billed for. "Cached" means previously processed inputs are reused, which makes them cheaper.

Long, difficult engineering tasks

Carbon is designed for sustained, agentic workflows, meaning cases where the AI plans and carries out multiple steps on its own. According to Google, internal teams ran the model for days at a time. It is said to push past dead ends and solve complex tasks, such as finding and fixing the root cause of subtle bugs in large codebases and low-level software. Google also says it improved coding, agentic and scientific tasks, full-stack web development and multimodal understanding (handling several data types such as images and text). On WebDev Arena, a comparison of web development results, Carbon gains 89 Elo points (a rating system built from head-to-head comparisons) and lands at #4 with 1769.

Cybersecurity across the whole defense process

Google sees its biggest leap in cyber defense: finding, verifying and fixing vulnerabilities. According to Google, the external benchmarks were run without guardrails.

Google claims the top score in all three. Exploiting vulnerabilities matters to defenders because it helps them decide which holes to close first. Internally, Carbon produces secure code more than twice as often as Argon, according to Google. The Chrome team reports that Carbon performed more than 10 times better on high-quality patches than another lab's large model (Google doesn't say which). The cloud security platform Wiz sees Carbon in the lead on web penetration-testing benchmarks.

Safety and alignment

Carbon uses the same layered safety approach as Argon: training, real-time misuse defenses, resistance to prompt injection (hidden instructions in content meant to mislead the model) and chain-of-thought monitoring. Google reports these results:

• Less cheating: Successful cheating on Terminal-Bench 4.0 drops to 0.0% and cheating attempts to 0.4%, with improved accuracy. In internal coding environments, severe misbehavior alerts fell by almost 80%.
• More honesty: Unsupported claims roughly halve compared with Argon. The model openly states methodological limits or blocked actions instead of inventing missing evidence.
• Asking when unsure: Instead of guessing, Carbon gathers sources and flags contradictions before it acts.
• Staying in bounds: Overreaching behavior drops by 26% and under-delivery by 29%.
• Prompt injection: 0.0% attack success on the Gray Swan benchmark. Argon scored 0.7%.

Our take

That a provider gives a frontier model to security professionals first shows how sensitive vulnerability-finding and exploitation capabilities are: the same skills help attackers too. The US government testing and the Fairwind Program are an attempt to manage that. For regular users and developers, though, it means waiting for now. All the figures come from the vendor. The only independent check so far is the Artificial Analysis index, and Google itself is the one announcing the #1 spot. Many safety numbers (alerts, unsupported claims) come from internal tests with no public methodology. Also open: the launch date, availability in products, and whether the launch price will last.