Last updated: November 11, 2026
n8n 2.48.0, a pre-release published November 10, 2026, switches n8n Assistant’s default model to Claude Opus 5.5, links Apps to Assistant chats, lets you select elements in the Apps preview and stops agent runs when Gateway credits run out. It also fixes agent memory, tool-call and Anthropic thinking bugs.
Contents
- What’s new at a glance
- Stable, beta and the 3.2 pre-release
- Set or change the Assistant model
- Apps and agents changes
- Plans and Assistant credits
- Limits and what to check
- Frequently asked questions
n8n is a workflow automation platform with a chat-based builder, n8n Assistant, and a Preview agents feature. Version 2.48 is a weekly minor release, but several changes land on those AI features. This guide draws on the official 2.48.0 release notes on GitHub and n8n’s own documentation.
What’s new at a glance
| Change in 2.48.0 | What it means for you |
|---|---|
| Default Assistant model bumped to Claude Opus 5.5 | Self-hosted instances that never set a model now use anthropic/claude-opus-5-5 instead of anthropic/claude-opus-4-8 |
| Link Apps to n8n Assistant chats | An App can be tied to the Assistant conversation that built or edits it |
| Select an element in Apps preview | You can point at a specific element in the preview instead of describing it in words |
| App bindings UI | A new interface for managing App bindings |
| Stop agent runs when the Gateway credits wallet is empty | Agents that use Gateway credits halt instead of continuing without funds |
| Microsoft Work IQ OAuth2 credential | A new credential type in core |
| Migration report updates | Archived workflows hidden by default, a loading message, and links from notification emails that open the report filtered to the recipient |
The release also lists AI-related fixes: Anthropic thinking stays valid in agents after prompt rewrites and web search, observational memory keeps running after a user stops a human-in-the-loop card, abandoned tool calls are settled when an Agents runtime turn starts, and Telegram agent channels now receive button clicks. The AI builder’s fetch-url tool now returns the raw body for JSON, XML, CSV and YAML, and Nano Banana image models no longer appear in Gemini chat model pickers (GitHub release).
One security-relevant fix applies the redaction policy to log streaming, telemetry and error reports as well.
Stable, beta and the 3.2 pre-release
n8n ships a new minor version most weeks. Its changelog says the stable version is for production and the beta is the most recent release, which may be unstable. When we checked on November 11, the changelog listed 2.47.4 as stable and 2.48.1 as beta. GitHub marks both 2.48.0 and 2.48.1 as pre-releases; 2.48.1 adds expression-engine and Edit Image node fixes.
A 3.2.0 pre-release for the upcoming 3.x line was published on GitHub the same day as 2.48.0. n8n publishes a v3.0 breaking changes page and a migration report for teams planning that move.
Set or change the Assistant model
According to n8n’s self-hosted setup guide, the N8N_INSTANCE_AI_MODEL variable uses a provider/model format and now defaults to anthropic/claude-opus-5-5. Supported providers are anthropic, openai and openrouter. The docs suggest anthropic/claude-opus-5-5 or openai/gpt-5.5 as starting points for hosted models.
- Upgrade a test instance to 2.48.x. Keep production on stable until 2.48 is promoted.
- Add your key in the instance’s AI settings, or set N8N_INSTANCE_AI_MODEL_API_KEY in your .env file. If that variable is missing, n8n falls back to ANTHROPIC_API_KEY, OPENAI_API_KEY or OPENROUTER_API_KEY.
- Leave N8N_INSTANCE_AI_MODEL unset to use Claude Opus 5.5, or set it explicitly, for example openai/gpt-5.5 or openrouter/deepseek/deepseek-v4-pro.
- If you route through a gateway with N8N_INSTANCE_AI_MODEL_URL and do not set a model, confirm the gateway offers Claude Opus 5.5. If it does not, set the model to one it offers, such as anthropic/claude-opus-4-8.
- Make sure a sandbox is enabled (N8N_INSTANCE_AI_SANDBOX_ENABLED=true). The bundled n8n-sandbox suits local testing; n8n recommends Daytona for production.
- Restart n8n, open the editor and confirm the Assistant appears and responds.
For a local model, point N8N_INSTANCE_AI_MODEL_URL at any OpenAI-compatible endpoint, such as http://localhost:1234/v1. Our Ollama new features guide covers one way to run such an endpoint.
Prompt pattern that saves iterations
The n8n Assistant documentation recommends stating the trigger, the apps involved, the data to read or write, what happens on success and failure, and whether the Assistant should ask before publishing. A compact template: “When [trigger], read [data] from [app], then [action] in [app]. On failure, notify [channel]. Ask me before publishing.”
Apps and agents changes
Agents arrived in n8n 2.40 as a Preview feature. Each agent has a model, plain-language instructions, tools such as n8n nodes, MCP servers and workflows, plus skills, sub-agents, knowledge files and memory (n8n changelog). Version 2.48 fixes several runtime problems: stale tool calls are settled at the start of a turn, and Anthropic thinking stays valid after a prompt rewrite or web search.
The new Gateway wallet check is worth noting for cost control. Gateway credits are a prepaid balance that pays for model calls made by workflows and agents. In 2.48, agent runs stop when that wallet is empty.
For Apps, the release adds linking to Assistant chats, element selection in the preview and an App bindings UI; the release notes give no further detail. For code-first comparisons, see our OpenAI Agents API guide and Browser Use 1.0 guide.
Plans and Assistant credits
| Setup | Assistant availability | How it is paid for |
|---|---|---|
| n8n Cloud Starter and Pro | Available | Monthly Assistant credit allowance, reset on the first of each month at midnight UTC; owners can buy extra credits that last 12 months |
| Free trial | Available | One-time allowance; top-ups not available |
| Self-hosted Community, Registered Community, Business | Available | Your own Anthropic, OpenAI or OpenRouter account |
| Enterprise | Not in the current availability list; docs describe limited enterprise readiness | Assistant credits are not available on Cloud Enterprise |
Source: n8n Assistant credits documentation. Assistant credits pay only for conversations with the Assistant. Running a workflow, even one the Assistant built, does not use them. Start a new conversation for unrelated work to save credits.
Limits and what to check
- n8n Assistant, agents and instance context are all labeled Preview. n8n says generated workflows can contain mistakes and should be reviewed before production use.
- Changing the default model can change cost and behavior on self-hosted instances that relied on the old default. Pin N8N_INSTANCE_AI_MODEL if you need predictable results.
- The Assistant uses your n8n user’s permissions and never receives credential secrets. Do not paste keys into the chat.
- n8n has not published benchmark results for Opus 5.5 inside the Assistant. For the model itself, see Anthropic’s own documentation; our Claude Fable 5.2 guide covers a related Anthropic release.
Frequently asked questions
What is new in n8n 2.48?
n8n 2.48.0, released November 10, 2026, makes Claude Opus 5.5 the default model for n8n Assistant, links Apps to Assistant chats, adds element selection in the Apps preview and an App bindings UI, stops agent runs when the Gateway credits wallet is empty, and fixes several agent and Anthropic thinking bugs.
Is n8n 2.48 a stable release?
No. GitHub marks 2.48.0 and the 2.48.1 follow-up as pre-releases. When this guide was written, n8n’s changelog listed 2.47.4 as the current stable version and 2.48.1 as the current beta, so test 2.48 before using it in production.
Which model does n8n Assistant use by default?
From n8n 2.48.0, the N8N_INSTANCE_AI_MODEL variable defaults to anthropic/claude-opus-5-5. Earlier versions default to anthropic/claude-opus-4-8. You can set the variable to another provider/model value, such as openai/gpt-5.5 or an OpenRouter model.
Does n8n Assistant cost extra?
On n8n Cloud Starter and Pro, n8n Assistant uses Assistant credits from a monthly allowance that resets on the first of each month, and owners can buy extra credits. Self-hosted n8n does not use Assistant credits; the assistant runs on your own model provider account, which bills you directly.
Can I run n8n Assistant with a local model?
Yes. Set N8N_INSTANCE_AI_MODEL_URL to a local or custom OpenAI-compatible endpoint, such as http://localhost:1234/v1, and add an API key only if the server needs one. If you route through a gateway without setting a model, make sure the gateway offers Claude Opus 5.5.
Sources
- n8n 2.48.0 release notes (GitHub)
- Set up n8n Assistant (n8n Docs)
- Use n8n Assistant (n8n Docs)
- Assistant credits (n8n Docs)
- n8n changelog (n8n Docs)