Claude Fable 5.2: what European teams need to know

Model choice at most European teams settled into a simple rule this autumn: Opus 5.5 for nearly everything, because it matched the pricier Fable 5.1 on most work. Claude Fable 5.2 arrived on 9 November 2026 and moves the top again, since Anthropic's own table puts it ahead of Opus 5.5 on hard coding, computer use and long research tasks.¹ The price went the other way, down 20 per cent from Fable 5.1, although the new model still costs double what Opus 5.5 does.²

Claude Fable 5.2 in brief, checked 9 November 2026

The sections below cover what Fable 5.2 is, what independent tests say, what it costs on each plan and where the limits sit for a business in Europe.

---

What is Claude Fable 5.2?

Claude Fable 5.2 is Anthropic's most capable model open to every customer, built for work that runs long: agent sessions lasting hours, multistep research and analysis carried through to a finished document.³ The same underlying model ships as Claude Mythos 5.2 for verified security and life science organisations, while Fable carries filters for cybersecurity and biology so that anyone can use it.¹ You find it in the Claude apps, in Claude Code and in the Claude API as claude-fable-5-2, and through Amazon Bedrock, Google Cloud and Microsoft Foundry.³ How the whole model family fits together is covered in what Claude AI is.

Anthropic does not present Fable 5.2 as the model for everything, and its own launch page tells teams to start with Opus 5.5 and step up only when a job needs it.¹ The dial for how much the model thinks is called effort, meaning a setting from low to max that trades depth against time and cost. In the apps and in Claude Code the default is medium, whereas the API defaults to high, one step above the medium that Opus 5.5 uses.¹⁴ Answers also arrive up to 30 per cent faster than with Fable 5.1.¹

---

How does Claude Fable 5.2 compare with Opus 5.5 and Fable 5.1?

Launch pages always carry a table where the new model wins, so read Anthropic's numbers the way you read a job application: true, selected and written by the candidate.¹

FrontierCode 2.0 measures hard coding tasks, and Fable 5.2 leads Opus 5.5 there by under four percentage points, 49.1 against 45.5 per cent.¹ The gap widens on the Remote Labor Index, a test built from real freelance projects, where Fable 5.2 completes 41.3 per cent and Opus 5.5 completes 21.3.¹ Not every row goes its way, because Google's Gemini 4 Carbon scores 53.9 per cent on AutomationBench, the test of everyday business workflows, against 51.4 for Fable 5.2.¹ Everyday office work is exactly where its lead is thinnest.

Claude Fable 5.2 in independent tests

References matter more than the application, and three independent test firms published results on launch day. Vals AI ranks Fable 5.2 first of 36 models on its combined index with 71.31 per cent, although Sonnet 5.5 follows at 70.85, which is inside the margin of error.⁵ Artificial Analysis puts it first with 65 points and adds the bill: at max effort it was the costliest of 73 models to test, at US$26.54 per task against US$10.04 for Opus 5.5.⁶ The weak spots sit in law and tax, where Vals places it 20th of 37 models on its legal index and 27th of 84 on its tax test.⁵

ARC Prize, a foundation that runs puzzles which are easy for people and hard for AI, measured 93.8 per cent on its second test and 81.2 per cent on its newest.⁷ Max effort backfired on that newest test, where the score fell from 81.2 per cent on xhigh to 65.7, because 28 of 55 games ran into the 12-hour limit.⁷

Claude Fable 5.2 as a colleague

Ethan Mollick, a Wharton professor with early access, handed Fable 5.2 whole projects instead of single questions, which makes his account the most practical one published on launch day.⁸ Six working games for a small light-up keyboard came out on the first try, and a 182-page research paper took under an hour of his own time. The study behind that paper was never registered in advance, as sound research practice asks, and no person had checked the results in full.⁸

A longer project produced over 450 3D worlds, until the agents wrote themselves a rule book that grew to 87,500 words and output fell to about a dozen worlds a day.⁸ His conclusion matters for a manager: the built-in reviewers catch errors, yet none asks whether the work is too long or dull, so that judgement stays with a person.⁸

Caution is the complaint from a second early tester, Matt Shumer, whose test version refused to ship a product change until a privacy policy update had gone live.⁹ In the same tests an agent sent 33 emails to an outside agency without his approval, so the model errs in both directions.⁹ Anthropic's prompting guide gives the remedy: the model applies approval rules exactly as written, so name the actions that need your sign-off.¹⁰

The release notes add two more habits to plan for, and both show up without any change to your prompts.⁴ On low and medium effort a turn sometimes ends with a list of what remains, while helper agents, which the model starts readily, can make up a large share of a session's cost. Someone still has to read the result before it goes to a client.

---

What does Claude Fable 5.2 cost?

Fable 5.2 is billed per token, meaning the word pieces an AI counts, and the Claude API charges US$8 per million tokens in and US$40 out, about €7 and €35.² That is 20 per cent below Fable 5.1 at US$10 and US$50, and exactly double Opus 5.5 at US$4 and US$20. Cached text, meaning instructions and files the model has already read, costs US$0.20 per million tokens on both models, so long sessions narrow the gap.²

The price per token says little about the price per task, because a stronger model sometimes needs fewer attempts to get a hard job right. Anthropic's own chart shows Fable 5.2 on high reaching 45.3 per cent on FrontierCode 2.0, level with Opus 5.5 at its top setting, for about half the cost per task.¹ The reverse holds on max, where Artificial Analysis needed 275 million output tokens to run its index against a median of 73 million.⁶

Claude Pro and Standard seats on Claude Team leave Fable outside the plan limits, so it runs on usage credits, meaning extra use billed by consumption, on top of €18 a month.¹¹ Max at €90, Ultra at €180 and Premium seats on Team count Fable inside the plan, where up to half of the weekly limit goes to it at no extra cost.¹¹ Launch week softens the line: until 16 November 2026, Pro and Standard seats also get Fable for up to half the weekly limit.¹² Anthropic states that Fable uses up limits at a higher rate than the other models, so run one real job before you promise it to a whole team.¹¹

Licence prices do not change with the model, and through satori-claude we buy and run the Claude seats for €490 once and €14 per user a month, excluding VAT. The full plan table in euros sits in how much Claude costs.

---

When should a European team pick Claude Fable 5.2 over Opus 5.5?

Opus 5.5 on medium stays the default for a whole team, because it costs US$4 and US$20 per million tokens and covers reports, contracts and everyday code. Anthropic gives the same advice and calls Opus 5.5 its strongest daily driver.¹

Fable 5.2 earns its price on jobs that run for hours without anyone watching, such as a migration across a company's entire code or research that ends in a finished report. Start it on high, since Anthropic's guide says most of the gain is already there, and go to xhigh only where you have measured a difference.⁴ Max is for measuring the ceiling, as the 12-hour timeouts in the ARC test showed.

Model choice does not have to be a guess, because a Claude consultant goes through the team's most common tasks and puts the right model on each. Most of it you manage alone: write down the five tasks the team spends most hours on and set Opus, Sonnet, Haiku or Fable next to each with the checklist below. Nobody should pay Fable prices to write an email.

Checklist: which Claude model for which job

• Emails, minutes and reports: Opus 5.5 on medium, or Sonnet 5.5 at half the price.
• Medium gets stuck on the same task: Opus 5.5 on high.
• Hours without supervision, or a job Opus on high has failed twice: Fable 5.2 on high.
• Personal data that must stay in the EU, or zero retention: Opus 5.5, because Fable 5.2 keeps prompts for 30 days on every route.
• Searches, logs and lookups: Sonnet 5.5 or Haiku 5.5.

Tip from satori.: Keep Opus 5.5 as the working model in Claude Code and let it consult Fable 5.2 as an advisor on the hard steps, a setup Anthropic describes in its launch post.¹ The Fable price then applies to a small share of the job.

---

Claude Fable 5.2, GDPR and the safety filters

Fable 5.2 requires 30 days of data retention so that Anthropic's systems can check for misuse, and zero data retention is unavailable unless Anthropic expressly authorises it.¹³ Opus 5.5 can run under zero retention, so a team that set that up for contracts or staff matters has to decide per workspace before switching.¹⁴ Write the 30 days into your record of processing, since GDPR Article 30 asks for the time limits for erasure, and check the data processing agreement before client files go through the model. The wider picture sits in our guide to Claude and GDPR.

EU routing for Fable 5.2 exists on one of the three routes we checked on 9 November 2026, and it comes with a catch. Anthropic's own API sends requests either globally or to the US only, and that holds for every current model.¹⁵ Amazon Bedrock lists Fable 5.2 in all eight of its European regions, but only through global routing, where any Amazon region worldwide may process the request.¹⁶ Google Cloud does offer its EU multi-region endpoint, yet it stores every prompt and response for up to 30 days and requires you to share them with Anthropic.¹⁷ Work that must stay in the EU therefore stays on Opus 5.5, Sonnet 5.5 or Haiku 5.5 through EU cross-region inference on Amazon Bedrock.

Cybersecurity questions that trip the filter are answered by the older Opus 4.8 and biology questions by Opus 5, so an IT team asking about a security flaw may hear from a different model.¹ Anthropic reports about 55 per cent fewer wrongful blocks than with Fable 5.1, a figure that comes from its own early testing.¹ Vals measured the other side: another model answered 1.88 per cent of its test requests.⁵

---

What changes in the Claude API with Claude Fable 5.2?

Teams with their own integrations against the Claude API face four changes that break code written for Fable 5.1:⁴

• Editing an earlier message, the system prompt or the tool list in a running conversation returns error 400 on every account.
• Only Fable 5.2 and Mythos 5.2 read the thinking steps Fable 5.2 has saved, so a switch to another model in the middle of a conversation loses them.
• The old computer control tool, computer_20251124, is rejected everywhere except on Amazon Bedrock.
• With the advisor tool, a Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor.

Moving from Opus 5.5 adds three more: the default effort rises from medium to high, the rate limits are separate and lower, and a workspace set to zero retention gets error 400.¹⁴ Search the code for tool_choice, computer_20251124 and effort before you change the model name to claude-fable-5-2, so the errors surface in testing and not at a client.

---

Summary: Claude Fable 5.2 is the specialist, Opus 5.5 stays the default

Opus 5.5 remains the sensible default for a European team, at half the token price, with zero retention available and an EU route through Amazon Bedrock. Fable 5.2 is the model you bring in for the jobs Opus fails, and the independent tests agree that it leads on long technical work while trailing on law and tax.⁵ Prices per seat in euros, VAT rules included, are on our page on Claude pricing in Europe.

Pick one task that Opus 5.5 got wrong last week, for example a spreadsheet analysis that needed three corrections, and run it again with Fable 5.2 on high. Add one line to the prompt: "Finish the whole task before you answer, then list what you are unsure about." Twenty minutes later you have two answers side by side, and that comparison says whether the double price buys your team anything.

---

Sources

Footnotes

1. Anthropic (2026). Introducing Claude Fable 5.2 and Claude Mythos 5.2, 9 November 2026. https://www.anthropic.com/claude-fable-and-mythos-5-2 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴

2. Anthropic (2026). Pricing, Claude Platform Docs, checked on 9 November 2026. https://platform.claude.com/docs/en/about-claude/pricing ↩ ↩² ↩³

3. Anthropic (2026). Claude Fable 5.2, model overview, Claude Platform Docs. https://platform.claude.com/docs/en/models/fable-5-2/overview ↩ ↩²

4. Anthropic (2026). What's new in Claude Fable 5.2, Claude Platform Docs. https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2 ↩ ↩² ↩³ ↩⁴

5. Vals AI (2026). Claude Fable 5.2 Benchmarks, Cost and Capabilities, checked on 9 November 2026. https://www.vals.ai/models/anthropic_claude-fable-5-2 ↩ ↩² ↩³ ↩⁴

6. Artificial Analysis (2026). Claude Fable 5.2 (Max, Default Fallback), checked on 9 November 2026. https://artificialanalysis.ai/models/claude-fable-5-2 ↩ ↩²

7. ARC Prize (2026). Claude Fable 5.2, ARC-AGI Results, 9 November 2026. https://arcprize.org/results/anthropic-claude-fable-5-2 ↩ ↩²

8. Ethan Mollick (2026). Fable 5.2 and Jagged AGI, One Useful Thing, 9 November 2026. https://www.oneusefulthing.org/p/fable-52-and-jagged-agi ↩ ↩² ↩³ ↩⁴

9. Matt Shumer (2026). My Fable 5.2 Review, Something Big, 9 November 2026. https://somethingbig.ai/fable-5-2-review ↩ ↩²

10. Anthropic (2026). Prompting Claude Fable 5.2, Claude Platform Docs. https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2 ↩

11. Anthropic (2026). Claude Fable models on your plan, Claude Help Center. https://support.claude.com/en/articles/15424964-claude-fable-models-on-your-plan ↩ ↩² ↩³

12. Anthropic (2026). Claude Fable 5.2 launch offers, Claude Help Center. https://support.claude.com/en/articles/17506494-claude-fable-5-2-launch-offers ↩

13. Anthropic (2026). API and data retention, Claude Platform Docs. https://platform.claude.com/docs/en/manage-claude/api-and-data-retention ↩

14. Anthropic (2026). Claude Fable 5.2 and Claude Mythos 5.2 migration guide, Claude Platform Docs. https://platform.claude.com/docs/en/models/fable-5-2/migration-guide ↩ ↩²

15. Anthropic (2026). Data residency, Claude Platform Docs. https://platform.claude.com/docs/en/manage-claude/data-residency ↩

16. Amazon Web Services (2026). Claude Fable 5.2, Amazon Bedrock model card, checked on 9 November 2026. https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-fable-5-2.html ↩

17. Google Cloud (2026). Anthropic's Claude Fable 5.2 on Google Cloud, Gemini Enterprise Agent Platform documentation, checked on 9 November 2026. https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/fable-5-2 ↩