Claude Fable 5.2 in n8n: when to use it in AI automation workflows

Published November 9, 2026 · Last updated November 9, 2026

Anthropic released Claude Fable 5.2 on November 9, 2026 (announcement). It replaces Fable 5.1 at a lower price and sits above Claude Opus 5.5 in Anthropic's lineup. Anthropic's own docs say to start with Opus 5.5 for most workloads and use Fable 5.2 for demanding reasoning and long-horizon agentic work (models overview). This page turns that advice into choices you make inside an n8n workflow: which node to use, which settings break, what a run costs, and which steps should stay on a cheaper model.

What changed on November 9

Sources: the Fable 5.2 model page, the pricing page and the announcement. At launch the models overview table showed no Bedrock ID, and the model page lists one now, which is why we say to check your region first.

How to call claude-fable-5-2 from n8n

You have two routes. The Anthropic nodes are quicker to wire into an AI Agent. The HTTP Request node gives you every API setting, including effort and fast mode, and works on any n8n version.

Option 1: the Anthropic nodes

1. Create an Anthropic credential in n8n with an API key from the Claude Console. Fable 5.2 needs 30-day data retention, so an organization on zero data retention gets a 400 error (retention rules).
2. Add an Anthropic Chat Model node under an AI Agent or Basic LLM Chain, or the Anthropic app node with Text > Message a Model for a single call. Both nodes load the model list live from Anthropic's /v1/models endpoint, so claude-fable-5-2 shows up in From List. You can also switch the field to ID and type it. Chat Model node versions 1 to 1.2 use a fixed list, so add a fresh node if yours is old (n8n docs).
3. Raise Maximum Number of Tokens. The defaults are 4,096 on the Chat Model node and 1,024 on the app node, and Fable 5.2 spends part of that on thinking before it writes the answer. Start at 16,000.
4. Do not add Sampling Temperature, Top K or Top P. Fable 5.2 returns a 400 error for them (sampling parameters). In n8n 2.30.0 and later the Chat Model node drops them for new Claude models on its own, but the app node sends whatever you add.
5. On the Chat Model node, leave Thinking Mode unset, or set it to Adaptive to choose an effort level. Disabled and Manual both fail on Fable 5.2, because thinking is always on and a thinking token budget is rejected. Note that the node's Adaptive mode defaults to medium effort, while the model's own default is high.
6. Raise the workflow timeout. On hard tasks one Fable 5.2 request can run for minutes, and the model page rates its latency as moderate.

Option 2: the HTTP Request node

Set the method to POST and the URL to https://api.anthropic.com/v1/messages. Under Authentication, pick Predefined Credential Type and your Anthropic credential, which adds the x-api-key header for you. Add the header anthropic-version with the value 2023-06-01 yourself, because the credential does not send it (n8n docs). Turn on Send Body, choose JSON, and send a body like this:

{
  "model": "claude-fable-5-2",
  "max_tokens": 16000,
  "output_config": { "effort": "medium" },
  "system": "You review inbound sales emails for an n8n workflow. Reply with JSON only.",
  "messages": [
    { "role": "user", "content": "{{ $json.emailBody }}" }
  ]
}

A few details that matter in a workflow:

• output_config.effort sets how hard the model thinks. The default on Fable 5.2 is high. For routine steps, try low or medium first, since thinking tokens bill as output (effort docs).
• Read stop_reason before you use the text. A safety classifier can decline a request with HTTP 200 and stop_reason: "refusal". Route that case with an IF node, or use the API's fallback option to retry on another Claude model (refusals and fallback).
• Add "speed": "fast" to the body for fast mode. It costs $16 / $80 per million tokens, so save it for steps where a person is waiting on the answer.
• For overnight jobs, the Message Batches endpoint takes 50% off. Fable 5.2 on Batch costs $4 / $20, the same per-token price as Opus 5.5 on a normal request.

When Claude Opus 5.5 is the cheaper choice

Fable 5.2 costs exactly twice as much as Opus 5.5 per input and output token. Cache reads are the exception: both models charge $0.20 per million cached tokens (model page). So the question in an n8n workflow is never "which model is better". It is "does this step fail often enough on Opus 5.5 to pay for the difference".

Keep the step on Opus 5.5, or a smaller model, when:

• It runs thousands of times a day: classification, routing, extraction, tagging, short replies.
• The output is easy to check with a rule, a schema or a Switch node, so mistakes get caught anyway.
• A person or a webhook caller is waiting. Fable 5.2 thinks longer by default.
• Opus 5.5 already passes your test set. Paying twice for the same pass rate buys nothing.

Move the step to Fable 5.2 when:

• An AI Agent node runs long, multi-tool tasks, and Opus 5.5 stops early, loops or loses the thread.
• The step produces a finished document, analysis or plan that a person would otherwise rewrite.
• You have tried Opus 5.5 at a higher effort and it still falls short on your examples.

Anthropic makes one cost claim worth knowing. On its FrontierCode coding benchmark at high effort, Fable 5.2 scores 45.3%, about level with Opus 5.5's best of 45.5%, at about half the cost per task (announcement). That is Anthropic's measurement on coding work, not ours, and not on n8n workflows. Test it on your own steps before you switch.

A simple pattern: put the model ID in a Set node at the start of the workflow, send each item to Opus 5.5 first, and use an IF node to retry only the failed or low-confidence items on Fable 5.2. You pay the higher price only on the hard cases.

Cost per run at $8 / $40

The token counts below are examples, not measurements. Output counts include thinking tokens, which bill as output. Replace them with the numbers from your own n8n execution logs. Prices are list prices with no caching or batch discount.

The math for the first row: 2,000 input tokens × $8 per million is $0.016, and 500 output tokens × $40 per million is $0.020, so $0.036 a run. The same run on Opus 5.5 is $0.008 + $0.010 = $0.018. Two levers change the picture. If most of a long agent run's input is a cached system prompt or document, those tokens cost $0.20 per million on either model, and the gap shrinks toward the output side. If Opus 5.5 needs a second attempt on one run in two, its real cost per finished task rises by half, and Fable 5.2 gets close.

AutomationBench: where Fable 5.2 leads and where Gemini does

AutomationBench tests business workflow tasks, which is the closest public benchmark to what an n8n team builds. These are Anthropic's published numbers from the launch table. Effort settings differ per cell, so read the footnotes on that page before you quote one.

Gemini 4 Carbon leads this row at 53.9%, and Fable 5.2 is second at 51.4%. It is one of the few rows in Anthropic's table where Fable 5.2 does not finish first. The bigger jump for Claude users is from Fable 5.1 (31.4%) and Opus 5.5 (42.5%). If your automations are mostly business workflow steps and you can use any provider, test Gemini 4 Carbon on the same items before you commit.

Artificial Analysis numbers: pending

As of November 9, 2026, Artificial Analysis has not published Intelligence Index, speed or price results for Fable 5.2. We will add them when they appear and will not estimate them before then.

Start from a workflow template

The fastest test is a workflow you already run. Swap the model in one step, run it on 20 real items, and compare the output and the cost in the execution log. If you need a starting point, browse the Claude workflow templates on AY N8N and change the model ID in the Anthropic node.

FAQ