Do I have to rebuild my n8n workflow when a new model launches?

No. In n8n the model is a sub-node attached to the Agent, so the trigger, the CRM node and the error branch do not change. You pick a different model in the dropdown and supply a credential. What can break is downstream parsing, so re-run your strictest output schema on real inputs before you switch the live workflow.

Can I use Gemini 4 Argon in n8n without a Google account?

On n8n Cloud, yes. n8n says Argon can run through Gateway credits, and the Google Gemini Chat Model node offers a Use Gateway credits option in its credential field, so you do not need a separate provider account or extra billing. On self-hosted you supply your own Google API key.

Why can the n8n Assistant not use Gemini directly?

n8n says the Assistant has no native Gemini option yet. On self-hosted you reach Gemini through OpenRouter or through Google's OpenAI-compatible endpoint. On n8n Cloud the Assistant runs on its own credits, and the workflows it builds can still use Argon through Gateway credits.

Gemini 4 Argon in n8n: swap the model, not the workflow

What shipped

Google announced Gemini 4 Argon on September 30, 2026 as its frontier model for long-horizon work: real software engineering, document-heavy knowledge work, and cyber defence (Google blog).

For automation the headline number is the token budget. Argon carries a 1M unified input and output limit through the Gemini API, with a new continuation protocol for long decodes (DeepMind). The output ceiling used to be 64K, so this is the part that changes what an agent can finish in one call.

Pricing is introductory: $2 per million input tokens and $10 per million output, cached input at 95% off. Google's own footnote puts standard pricing at $4 and $20 once the introductory period expires on November 30, 2026.

Then on October 13 n8n published its own post. Argon is a model option for n8n Agents and in the standalone Google Gemini node from launch, and on n8n Cloud you can run it on Gateway credits instead of wiring up your own Google account (n8n).

What this means for a small operator

Your lead scorer, document parser or support triage agent can get a new brain without you touching the nodes around it.

That is the whole business angle. In n8n the model is a sub-node hanging off the agent. The trigger, the CRM node, the error branch and the output parser have no idea which model answered. The contract between them is the output schema, not the vendor.

So the question is not whether you rebuild. It is whether your workflow is built so the model is a swappable part. If your prompt is pasted inside a Code node, if your parsing depends on one model's formatting quirks, or if you have no way to compare two runs on the same input, then every launch costs you a week. That is a design problem, not a Google problem.

Decision table: what Argon actually touches

Part of your stack	Does Argon touch it	What to do
The model sub-node on an Agent	Yes	Pick Argon in the dropdown, leave the rest of the canvas alone
Credentials	Only on self-hosted	On n8n Cloud, select Use Gateway credits in the node's credential field (n8n docs)
Trigger, CRM node, Slack, error branch	No	Nothing
Your output parser	Maybe	Re-run your strictest schema and watch for format drift
The n8n Assistant	Not yet	No native Gemini option, so route through OpenRouter or Google's OpenAI-compatible endpoint
Your cost model	Yes	Re-price it before the introductory window closes on November 30

The swap, about twenty minutes

Duplicate the live workflow and disable the copy's trigger.
On the copy, point the Agent's model sub-node at Argon.
Collect a set of real inputs where you already know the right answer. Yesterday's leads work.
Run both workflows on the same inputs and diff the output fields, not the prose.
Leave the old model configured in a disabled node so rollback is a click.

The third item is the one everybody skips, and it is the only one that tells you anything.

Limitations

Benchmarks are not your workload. Argon ranks first of 48 models on the Vals Index at 68.9% (Vals), and third-party testing put its hallucination rate at 15%, the lowest among leading models (compiled analysis). Neither figure tells you whether it reads your quotes correctly.

Access is staged. The rollout went to developers on the paid Gemini API tier, to Ultra subscribers and to enterprises, and a compiled analysis notes that Google had not published an official API model ID in the days after launch. Check what your own account can call before you plan around it.

I have not tested Argon inside an n8n agent. Everything above comes from the documentation and the launch sources, and the time estimate is mine.

The pricing switch is a real risk. A workflow that looks cheap this month gets re-priced when the introductory window closes, and long output is where that bites.

My take

This launch is worth about twenty minutes of your attention, not a project. Pick your single most expensive agent, run the copy, keep the receipts. Being able to swap a model in an afternoon is worth more than any one model, and it is the thing you can actually build this week. More on how I wire these up on the blog.