Skip to main content
Yarin AmrusiAutomation and systems development

Claude Fable 5.2: What to Check Before Switching ERP Scripts

Anthropic released Claude Fable 5.2 on November 9, 2026. Which changes can break existing ERP scripts, and how to test the switch without risking business data.

Published: Yarin Amrusi

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. On its announcements page, the company says the models set a new bar in coding and finance. Some businesses use Claude to build scripts that pull data from SAP Business One or Priority (an ERP system common in Israel), or integrations that move orders and invoices between systems. They may find that swapping the model name in their code breaks things that used to work.

Changes that can break existing code

A developer who went through the migration guide posted a summary on DEV Community. It is a personal write-up, not Anthropic's official guide. It lists these changes:

  • 30-day data retention. Fable 5.2 requires it. An API organization without that setting gets a 400 error on every request. For a business that sends customer data to the model, this is a management decision, not only a technical setting.
  • Editing conversation history. Code that changes the system prompt, the tools or earlier messages before sending the model's thinking block back now gets a 400 error. On 5.1 this check applied only to accounts created on or after August 31, 2026. On 5.2 it applies to every account.
  • Fallback models. If a fallback moves the conversation to 5.1 or Opus 5.5, the API drops Fable 5.2's thinking blocks. The request still succeeds, but cost and response time go up.
  • Old tools. The computer-use tool computer_20251124 returns an error, except on Bedrock. When Fable 5.2 is the executor, the Advisor tool must be Fable 5.2 or Mythos 5.2.

One more change breaks nothing, but it affects the bill. According to Digital Applied, the default effort level is Medium in the Claude apps and Claude Code, and High on the API. A good result in the chat window says nothing about the cost or behavior of the same task called from code.

Anthropic points routine work elsewhere

In its launch post, Anthropic describes Fable and Mythos as models for "the most ambitious, open-ended technical work." For most coding and knowledge tasks, it recommends starting with Claude Opus 5.5, then moving up to Fable or down to Sonnet or Haiku 5.5 as needed.

Its own numbers explain why. On FrontierCode 2.0, Fable at High effort scored 45.3%, against 45.5% for Opus 5.5's best result, at what Anthropic calls "roughly half the cost per task." At Xhigh effort it reached 49.1%. Effort level is a setting in each call to the model, and the cost of a task depends on it.

On AutomationBench, which tests business workflows, Anthropic's table shows a jump to 51.4% for Fable 5.2, up from 31.4% for 5.1. Opus 5.5 scored 42.5% and GPT-6 Astra 41.4%. Gemini 4 Carbon still leads the same table at 53.9%. The benchmark's publisher and methodology could not be found, so these figures come only from the vendor.

A query that runs can still be wrong

The most telling example in the launch post comes from Oracle. Its test combined country debt summaries, repayment history and regional benchmarks. Fable 5.2 wrote correct SQL. Two other models wrote queries that ran without a single error but returned values "5–8x the expected amounts." The vendor chose this example, and it is not an independent test.

The lesson applies to any report pulled from an ERP. In a collections report or a list of supplier balances, the dangerous error is probably not the one that crashes the script. It is the one that returns a plausible number nobody stops to question.

An independent test by SonarSource found better code quality. Fable 5.2 passes 90.26% of test-backed tasks, against 87.68% for Fable 5 and 5.1. Issue density fell 22.0%, from 17.44 to 13.60 per thousand lines of code. Four blocker vulnerabilities remain, and the remaining bugs cluster in concurrency, meaning code that runs in parallel. SonarSource advises verifying every change. That probably matters most for integrations that sync orders and stock across several processes at once.

How to test the switch safely

  1. Search the code for history edits. Find every place that changes the system prompt, the tool list or earlier messages before resending. The developers who summarized the guide recommend this step.
  2. Check the data retention policy. Ask whoever manages the API account whether the organization is set to 30-day retention, and whether that is acceptable when customer data passes through it.
  3. Switch models in a test environment only. The API model ID is claude-fable-5-2. Set the effort level explicitly in code, and run the same tasks on Opus 5.5 for comparison.
  4. Check results against known numbers. Run the reports on a month that is already closed and whose totals the accounts team has verified, then compare them figure by figure.
  5. Watch the fallback. Log when requests move to the fallback model, and measure cost and response time before and after the switch.

When not to rush

Automations that read emails and files from outside the business are more exposed. Notebookcheck reported figures from the model's system card on prompt injection, meaning hidden instructions inside an email or file. These attacks succeeded against Fable 5.2 in 4.09% of cases, compared with 1.02% for Fable 5.1 and 0.96% for Opus 5.5. Most of the gap comes from the fallback model: Opus 4.8 answered 53% of the attempts and was hijacked 5.17% of the time. When Fable answered itself, it was hijacked 2.05% of the time. An agent that reads supplier invoices from email and can open ERP documents needs a person to approve every action.

The sources say nothing about pricing, Hebrew-language support or availability in Israel. Businesses there need to check all of that before committing.

Switching makes little sense when existing scripts work and the tasks are routine, such as pulling a report or updating an order status. Anthropic itself sends that kind of work to Opus 5.5. Nor does the switch suit a business that cannot approve 30-day data retention, or an automation that acts on outside emails without human approval. Fable 5.2 is worth testing when the current model fails at a complex coding task and someone can review every change before it goes live.

A short checklist

  • Is the API organization set to 30-day data retention, and is that acceptable for customer data?
  • Does the code change the system prompt, tools or earlier messages before sending a request back to the model?
  • Is the effort level set explicitly in code rather than left to the default?
  • Is there a report for a closed period with verified totals to compare against?
  • Which model does the fallback switch to, and what does that do to cost and response time?
  • Does a person approve any ERP action that is based on an outside email or file?

The home page has screenshots of systems I built. If you have a similar process that takes up time, you can tell me about it in a short conversation.

Frequently asked questions

How much does Fable 5.2 cost in an existing integration?

The sources do not give a price. Anthropic says that on FrontierCode 2.0 at High effort, Fable 5.2 came very close to Opus 5.5's best result at roughly half the cost per task. Actual cost depends on the effort level set in code, and it rises when requests move to a fallback model. It is worth measuring on the business's own tasks.

Is trying the model in the Claude chat app enough before connecting it to code?

It gives a first impression, but it is not a full test. According to Digital Applied, the default effort level is Medium in the Claude apps and High on the API. A result in the chat app therefore says nothing about the cost or behavior of the same task when it runs from a script or an ERP integration.

What does the 400 error in the migration guide mean?

It means the API has rejected the request. According to the DEV Community summary, Fable 5.2 returns it on every request if the organization is not set to 30-day data retention. It also returns it whenever the code edits the conversation history before sending the thinking block back. Check the integration's error logs right after the switch.

Sources