The n8n blog announced that Google's Gemini 4 Carbon is now available inside n8n. Google describes it as its strongest model so far, with gains in coding, long-running agentic work and cybersecurity. All benchmark figures below are reported by Google and relayed by n8n; we have not verified them independently.
What changed in n8n
In an AI Agent node, you add the Google Gemini Chat Model and pick Gemini 4 Carbon from the model list. The model also appears in the standalone Google Gemini node. As with any other model, you then attach the tools the agent may use and add memory if it needs context.
Access depends on how you run n8n. On n8n Cloud, Carbon is available through Gateway credits, so there is no separate provider account or extra billing. On self-hosted installations you add your own Gemini API key.
The figures Google published
- 53.9% on AutomationBench, a benchmark for automation work, which Google says is the top result.
- Security tasks: 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (checking which ones are exploitable) and 71.0% on CWE-bench (patching).
- On Gray Swan's prompt injection benchmark, a 0.0% attack success rate, compared with 0.7% for the earlier Gemini 4 Argon.
n8n itself adds a caveat: no model is fully resistant to malicious attacks at scale, so existing security protocols should stay in place.
Long tasks and untrusted input
Google says internal runs on complex engineering problems have lasted for days, with the model working past dead ends. n8n reads that as agents being able to get through more steps and tool calls before losing the thread.
The post's example is a security alert arriving by webhook. The agent reads the alert, gathers matching logs, reviews the known-issues list, assigns a priority, and sends a summary to the on-call engineer in Slack. The alert text is untrusted input, which is where the prompt injection result matters most.
Pricing and mixing models
Carbon is debuting with an introductory rate that matches Argon's: $2 for every million input tokens and $10 for every million output tokens. When the introductory period ends on November 30, 2026, the rates rise to $4 and $20.
n8n suggests that teams already on Argon need not move everything. Carbon can take the steps that need heavy reasoning, while Argon or a lighter model handles simple work such as classification or formatting.
What the source does not say
The post gives no detail on Gateway credit pricing for Carbon beyond saying there is no extra billing, and it offers no independent test results. The benchmark claims come from Google's own announcement.