n8n 2.44 is stable: prompt caching for Claude has to be switched on per node

27 October 2026 Michael Engel Automation Costs

The AI features from n8n 2.42 to 2.44 are now in the stable line, but caching for Claude only saves money once you switch it on.

Source: n8n@2.44.7 (n8n-io/n8n on GitHub), 26 October 2026

Note: I use Claude and Claude Code for my own work. Anthropic does not pay me, and the links are not affiliate links.

Technical drawing: a workflow of connected nodes with a highlighted agent node and its cache stack, below it a switch flipped from OFF to ON, on the left the label STABLE.

What happened: 2.44 has been the stable line of the automation platform n8n since 19 October (2.44.4). On 26 October, 2.44.7 followed with bug fixes only. On GitHub it is “Latest”, and on npm and Docker it carries the stable tag, which according to the n8n documentation is the version for production use. The stable line brings these AI features (release notes):

2.42 (29 Sept): prompt caching in the Anthropic node (“Message a Model” operation); the Agents feature is on by default.
2.43 (6 Oct): Browser Use in the n8n Assistant for all users, which admins can turn off with an instance setting; an optional budget limit for agents.
2.44 (13 Oct): credentials for Google Vertex AI; the Microsoft Foundry chat model node (previously “Azure AI Foundry”) can now talk to Claude models deployed there.

What this means: With prompt caching, Anthropic stores the fixed beginning of a request, such as a long system prompt (the standing instructions to the model). According to Anthropic, reading from the cache usually costs a tenth of the normal input price, and a twentieth for Opus 5.5 and Sonnet 5.5. Writing costs more: 1.25 times the input price for a 5-minute lifetime, twice the price for one hour. In n8n the option is off by default, both in the Anthropic node and in the Anthropic Chat Model sub-node used with the AI Agent. In your workflows, updating alone saves nothing.

What you can do:

Test first: Install 2.44.x on a test instance before production.
Switch caching on where it pays: In Claude nodes with long, fixed instructions, set “Prompt Caching” to “5 Minutes”. Use “1 Hour” only if the same workflow runs less often than every five minutes but within the hour.
Order the prompt: The fixed part goes first and stays identical character for character; dates and other variables go at the end. Below 512 tokens (the billing units for text; the threshold for Opus, Sonnet and Haiku 5.5), Anthropic does not cache at all, and no error is shown.
Check the bill: Since 2.42, the Anthropic node includes cache tokens in its reported token usage. If no cache read tokens show up, you are only paying the surcharge for writing. Compare your Anthropic bill before and after the change.

Context: Prices and features are vendor statements (Anthropic, n8n).

Bringing AI into your own team? That’s what my workshops are for. See the workshops