---
title: "Claude Fable 5 eCommerce and Claude Commerce: 2026 Guide"
description: "Claude Fable 5, 5.1 and 5.2 for eCommerce teams: prices, the June export-control pause, API changes, refusals and running Claude Commerce agents on Fable."
url: "https://no7software.co.uk/blog/claude-fable-5-ecommerce-engineering"
last_modified: "2026-11-11"
---

# Claude Fable 5 eCommerce and Claude Commerce: 2026 Guide

## Key takeaways

- Claude Fable 5.2 (9 November 2026) is the current Fable model at $8/$40 per million tokens; Fable 5 and 5.1 are legacy at $10/$50.
- Fable 5 was suspended from 12 June to 1 July 2026 after a US export control directive, not a technical fault.
- Fable rejects budget_tokens, disabled thinking and, on 5.1 and 5.2, forced tool_choice; use the effort parameter instead.
- Every call needs a stop_reason refusal check and a fallback model, because the classifiers also refuse some benign requests.
- Claude Commerce Agents, Anthropic's open-source shopping and merchant agent blueprint, defaults to Sonnet 5 and Opus 5; running it on Fable 5.2 needs thinking turned on and no forced tool_choice.

Claude Fable is Anthropic's top model tier for long-horizon reasoning and agentic work. Claude Fable 5 launched on 9 June 2026 at $10 per million input tokens and $50 per million output tokens, Claude Fable 5.1 followed on 1 September 2026 at the same price, and Claude Fable 5.2 arrived on 9 November 2026 at a lower $8 and $40. For eCommerce engineering teams, the practical questions are which version to call, what the API change breaks in existing code, how to handle the model's safety refusals, and when a cheaper model does the same job. We also cover Claude Commerce, Anthropic's open-source blueprint for shopping and merchant agents, and what it takes to run it on Fable. Every figure below comes from Anthropic's own model pages and announcements.

## Which Claude Fable version should you use today?

Use Claude Fable 5.2 (`claude-fable-5-2`) for new work. Anthropic's [Fable 5.2 overview](https://platform.claude.com/docs/en/models/fable-5-2/overview) marks it as the latest Fable model and lists Fable 5 and Fable 5.1 as legacy models that are still available. Fable 5.2 also costs less than the versions before it. Anthropic's [models overview](https://platform.claude.com/docs/en/models/overview) recommends starting most workloads on Claude Opus 5.5 and moving to Fable 5.2 when your evaluations on Opus 5.5 at higher effort still fall short.

| Model | Released | Price per million tokens (input / output) | Default effort | Knowledge cutoff |
| --- | --- | --- | --- | --- |
| Claude Fable 5.2 | 9 November 2026 | $8 / $40 | high | June 2026 |
| Claude Fable 5.1 (legacy) | 1 September 2026 | $10 / $50 | high | June 2026 |
| Claude Fable 5 (legacy) | 9 June 2026 | $10 / $50 | high | January 2026 |
| Claude Opus 5.5 | Current | $4 / $20 | medium | June 2026 |
| Claude Sonnet 5.5 | Current | $2 / $10 | high | June 2026 |
| Claude Haiku 5.5 | Current | From $0.10 / $0.50 | medium | June 2026 |

All six models have a 1 million token context window and up to 128,000 output tokens per request. Prices and dates are from each model's page on Anthropic's platform documentation; Haiku 5.5's lower price applies to prompts up to 100,000 tokens.

## Why was Claude Fable 5 suspended after launch?

Because of a US export control directive, not a technical fault. Anthropic's [redeployment post](https://www.anthropic.com/news/redeploying-fable-5) sets out the timeline:

- **9 June 2026:** Anthropic released Claude Fable 5 and Claude Mythos 5. Both share the same underlying model; Fable 5 carries strong safeguards for general use, while Mythos 5 has fewer safeguards and went only to a small group of partners for defensive cybersecurity.
- **12 June 2026:** the US government applied export controls to both models, requiring Anthropic to block access by foreign nationals. With no reliable way to verify nationality in real time, Anthropic suspended both models for every user, as its [statement on the directive](https://www.anthropic.com/news/fable-mythos-access) explains. Access to Anthropic's other models was not affected.
- **30 June 2026:** Anthropic announced the export controls had been lifted.
- **1 July 2026:** Fable 5 became available again globally on the Claude Platform, Claude.ai, Claude Code and Claude Cowork, with updated cybersecurity safeguards.

For a business, the lesson is about dependency rather than this one model: any feature built on a single frontier model needs a route to a second model if access stops. The fallback pattern later in this guide covers that.

## What changes in your API code when you move to Fable?

Fable models reject several request settings that older Claude models accepted. Anthropic's [What's new in Claude Fable 5.2](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2) page lists them:

- **Thinking is always on.** Sending `thinking: {type: "enabled", budget_tokens: N}` or `thinking: {type: "disabled"}` returns a 400 error. Omit `thinking`, or send `{type: "adaptive"}`.
- **Effort replaces thinking budgets.** Control depth with `output_config.effort`: `low`, `medium`, `high`, `xhigh` or `max`. The default is `high`, and Anthropic advises starting there and stepping up to `xhigh` only for the hardest long-running work where you have measured a gain.
- **Forced tool use is gone on 5.1 and 5.2.** `tool_choice` of `any` or a named tool returns a 400. Keep `tool_choice: auto` and set `strict: true` on the tool for schema-valid arguments, or use structured outputs.
- **The older computer use tool** `computer_20251124` returns an error on Fable 5.2 on every platform except Amazon Bedrock.
- **Thinking blocks belong to the model that wrote them.** Only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks, and editing earlier turns of a conversation invalidates them, so build your agent history as append-only.

Before switching an existing integration, search your code for `budget_tokens`, `temperature`, `top_p`, `top_k` and forced `tool_choice`, and remove them for Fable requests.

## How should a store handle Fable's safety refusals?

Check `stop_reason` on every response before you read the content. Fable models run safety classifiers, and a declined request still returns HTTP 200 with `stop_reason: "refusal"` and a `stop_details` object naming the policy area that fired. A refusal looks like this:

```
{
  "stop_reason": "refusal",
  "stop_details": {
    "type": "refusal",
    "category": "cyber",
    "explanation": "..."
  }
}
```

Anthropic says it deliberately set Fable 5's classifiers with a wider safety margin than any earlier launch, which means some reasonable requests are refused too. For a production agent, plan for that rather than treating it as an edge case. Anthropic's [refusals and fallback guide](https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback) offers three options: server-side fallback, the SDK's client-side fallback middleware, or your own retry. Server-side fallback is in beta on the Claude API; with it, `fallbacks: "default"` retries a declined request on the model Anthropic recommends for that category:

```
POST /v1/messages
anthropic-beta: server-side-fallback-2026-07-01

{
  "model": "claude-fable-5-2",
  "max_tokens": 16000,
  "fallbacks": "default",
  "messages": [{ "role": "user", "content": "..." }]
}
```

For Fable 5.2, the permitted fallback targets are Claude Opus 4.8 and Claude Opus 5. Some categories, such as `reasoning_extraction`, have no fallback and the refusal stands, so your code still needs a path that tells the user or queues the task for a person.

## What does a Fable run cost compared with Opus and Sonnet?

Work it out per task, not per token. Take a codebase task that sends 200,000 input tokens and returns 20,000 output tokens, at Anthropic's list prices:

| Model | Input cost | Output cost | Total per run |
| --- | --- | --- | --- |
| Claude Fable 5 or 5.1 | $2.00 | $1.00 | $3.00 |
| Claude Fable 5.2 | $1.60 | $0.80 | $2.40 |
| Claude Opus 5.5 | $0.80 | $0.40 | $1.20 |
| Claude Sonnet 5.5 | $0.40 | $0.20 | $0.60 |

Three things move the real number. Prompt caching cuts repeated input sharply: on Fable 5.2, cache reads cost 2.5% of the base input price. The Batch API halves the price for work that can wait: $4 input and $20 output per million tokens on Fable 5.2. And the thinking the model does is billed whether or not you display it, so a higher effort level raises output cost. The cheaper model only wins if it finishes the task: a run that needs a second attempt costs twice.

## Where does Fable earn its price in eCommerce engineering?

On long, multi-step work where a weaker model would need several attempts or a person to stitch the pieces together. In a commerce codebase that usually means:

- **Platform migrations:** mapping a Magento or Salesforce Commerce Cloud schema onto Shopify products, metafields and metaobjects, with the full export in context.
- **Shopify Functions under pressure:** Shopify's [Functions documentation](https://shopify.dev/docs/apps/build/functions/optimize-instruction-counts) caps a function at 11 million WebAssembly instructions for carts of up to 200 line items. Profiling and refactoring a function that runs close to that limit is reasoning-heavy work.
- **Integration refactors:** changes that ripple through an ERP, a PIM and order webhooks at the same time, where the model must hold all three in view.

Route everything else down. Theme copy, CSS tweaks, product description drafts and catalogue tagging belong on Sonnet 5.5 or Haiku 5.5, and general coding on Opus 5.5. A senior engineer still reviews whatever any of them produces. For how we set up Claude Code against Shopify, see our [Shopify AI Toolkit with Claude Code guide](https://no7software.co.uk/blog/shopify-ai-toolkit-claude-code-production).

## Is Claude Commerce powered by Fable 5?

No. "Claude Commerce" usually means Claude Commerce Agents, an open-source reference blueprint Anthropic published on GitHub in September 2026 ([anthropics/commerce-agents](https://github.com/anthropics/commerce-agents), Apache 2.0). Its shopping agent defaults to Claude Sonnet 5 and its merchant agent to Claude Opus 5, not Fable. The model is a single config string, so you can point either agent at Fable 5.2.

The blueprint is code you run yourself. It is not a hosted Anthropic store and not a new model. Its README says nothing in it places an order, charges a card or changes a live listing: checkout hands the cart to your own site, and every merchant write waits until a person approves it. The repository is also marked as not maintained. Shopify has published its own implementation, [Shopify/claude-for-commerce-examples](https://github.com/Shopify/claude-for-commerce-examples), with a storefront shopping agent over UCP and a merchant agent over the Admin API.

| Agent | What it does | Default model |
| --- | --- | --- |
| Shopping agent | Searches, compares, fills the cart, answers order and policy questions | `claude-sonnet-5` |
| Merchant agent | Explains performance, maintains listings, prices and promotes, drafts campaigns; every write is staged for approval | `claude-opus-5` |
| Memory extraction | Saves what a customer or merchant tells the agent after each turn | `claude-haiku-4-5-20251001` |

Defaults are from the repository's [deployment guide](https://github.com/anthropics/commerce-agents/blob/main/docs/deployment.md). Moving the blueprint to Fable 5.2 takes more than a new model string, because two of its defaults run into the API changes listed above:

- **Thinking is off by default.** With `thinking_effort` left at `None`, the turn loop sends `thinking: {type: "disabled"}`, which Fable rejects with a 400. Set `thinking_effort` to `"high"`. The default `max_tokens` of 2048 has to cover the thinking and the reply together, so raise it too.
- **The Messages API runtime forces a tool on grounded turns.** When a grounding rule fires, for example on an order question, the first round is pinned with `tool_choice: {type: "tool"}`, which Fable 5.1 and 5.2 reject. The Agent SDK runtime runs those grounding reads on the host instead, so it avoids the problem. On the Messages API runtime you would need to change the orchestrator.

```
from merchant_agent import MerchantAgentConfig

config = MerchantAgentConfig(
    brand_name="Your Store",
    model="claude-fable-5-2",
    thinking_effort="high",
    max_tokens=16000,
)
```

For most stores the defaults are the better starting point. Keep Sonnet 5 on the shopping agent, where requests are short and frequent. If you want to try Fable, start with the merchant agent's analysis delegate (`analysis_model`), where long reasoning over sales data is most likely to justify the price, and measure it against Opus 5 on your own data.

## Can Fable power a customer-facing shopping agent?

It can, but it is rarely the right default. A shopping agent answers many short requests, which is where cost per request and refusal handling matter most, and both favour a cheaper model with Fable kept for the hard cases. Connect the agent to your catalogue and cart through the Model Context Protocol, and for checkout use the commerce protocols compared in our [ACP, UCP and MCP guide](https://no7software.co.uk/blog/agentic-commerce-protocols-acp-mcp-ucp-explained). Whichever model you choose, the refusal check and fallback above are part of the agent, not an extra.

## What to do next

Search your Claude integrations for the settings Fable rejects, add a `stop_reason` check and a fallback to every call, and run your own tasks on Opus 5.5 before paying for Fable 5.2. If you are still on Fable 5 or 5.1, moving to 5.2 lowers the per-token price by a fifth, so test the breaking changes above and switch. If you want help building or reviewing an AI integration for your store, our [Shopify development team](https://no7software.co.uk/shopify-development) can scope it.

## Frequently Asked Questions

The questions buyers and engineers ask us most about this topic.

### Which Claude Fable model should I use in November 2026?

Claude Fable 5.2 (claude-fable-5-2), released on 9 November 2026. Anthropic lists it as the latest Fable model and marks Fable 5 and Fable 5.1 as legacy. It costs $8 per million input tokens and $40 per million output tokens, against $10 and $50 for Fable 5 and 5.1. Anthropic also recommends starting most workloads on Claude Opus 5.5 ($4 and $20) and moving to Fable 5.2 only when evaluations on Opus 5.5 at higher effort fall short.

### Why was Claude Fable 5 unavailable in June 2026?

On 12 June 2026 the US government applied export controls to Claude Fable 5 and Claude Mythos 5, requiring Anthropic to block access by foreign nationals. Because it could not verify nationality in real time, Anthropic suspended both models for all users. The controls were lifted on 30 June, and Fable 5 was available globally again from 1 July 2026 with updated cybersecurity safeguards.

### How do I handle safety refusals from Claude Fable in production?

Check stop_reason on every response. A declined request returns HTTP 200 with stop_reason set to refusal and a stop_details object naming the policy area. Anthropic offers server-side fallback (beta on the Claude API), SDK fallback middleware, or your own retry; on Fable 5.2 the permitted fallback models are Claude Opus 4.8 and Claude Opus 5. Some categories have no fallback, so your code also needs a path that informs the user or hands the task to a person.

### Is Claude Commerce powered by Claude Fable 5?

No. Claude Commerce Agents, the open-source blueprint Anthropic published on GitHub in September 2026, defaults to Claude Sonnet 5 for its shopping agent and Claude Opus 5 for its merchant agent. You can set the model to claude-fable-5-2, but you also need to turn thinking on with thinking_effort and avoid the forced tool_choice the Messages API runtime sends on grounded turns, because Fable 5.2 rejects both.
