Anthropic released Claude Fable 5.2 on 9 November 2026, priced below its predecessor. If you run Fable 5.1, you'll want to switch. If you don't, the more useful question is why not.
Claude Fable 5.2 for business costs $8 per million input tokens and $40 per million output tokens, 20 percent less than Fable 5.1 but still double the price of Claude Opus 5.5. Anthropic's own documentation says to start with Opus and only move to Fable when your tests on Opus fall short. There's also a condition that weighs more for a Dutch company than the price: Fable models require Anthropic to keep your data for 30 days.
What changed with Claude Fable 5.2?
Claude Fable 5.2 is Anthropic's new top model for heavy reasoning and long agent tasks, released on 9 November 2026 as claude-fable-5-2. According to Anthropic's overview page it has a 1 million token context window and up to 128,000 output tokens, and adaptive thinking is always on. The effort parameter, high by default, sets how deeply the model thinks.
The price cut is the most visible change. A second model, Claude Mythos 5.2, has the same capabilities and the same price, but only organisations that have passed one of Anthropic's verification programmes can use it. For most companies, the choice is Fable.
| Model | Input / output per 1M tokens | 30-day retention required |
|---|---|---|
| Claude Fable 5.2 | $8 / $40 | Yes |
| Claude Fable 5.1 | $10 / $50 | Yes |
| Claude Opus 5.5 | $4 / $20 | No, zero data retention possible |
Prices are from Anthropic's pricing page. Batch processing on Fable 5.2 costs $4 / $20, fast mode $16 / $80.
What does Fable 5.2 cost compared with Opus 5.5?
Claude Fable 5.2 costs twice as much per token as Claude Opus 5.5. At 100 million tokens a month, three quarters of it input, with no caching or batch discount, you pay roughly:
- Opus 5.5: 75 × $4 + 25 × $20 = $800
- Fable 5.2: 75 × $8 + 25 × $40 = $1,600
- Fable 5.1: 75 × $10 + 25 × $50 = $2,000
If you're already on Fable 5.1, switching saves $400 a month at this volume. If you're on Opus, Fable costs $800 a month more. That's not necessarily a problem: if Fable gets a task right first time that Opus gets wrong one time in five, the higher token price can work out cheaper per finished task. The point is that you have to measure it. How is in our piece on cost per task.
Anthropic itself writes: "For most workloads, start with the latest Claude Opus model", and move to Fable 5.2 "when your evals on a Claude Opus model at higher effort still don't meet your requirements". A vendor pointing you at its cheaper model deserves a hearing. Without your own test set, though, you can't tell whether your work falls under those "most workloads"; building one is covered in evaluating AI benchmarks.
What does 30-day retention mean for GDPR?
The retention requirement means every prompt and response sent to Claude Fable 5.2 stays with Anthropic for thirty days, and zero data retention isn't available for this model. Anthropic calls Fable 5.2, Fable 5.1, Fable 5 and the Mythos variants "Covered Models". According to the data retention page these models "require 30-day data retention", so ZDR isn't available "unless expressly authorized by Anthropic". It isn't new in 5.2, but it still applies.
For a company with a zero data retention agreement on Opus, that means Fable won't run in the same workspace. Anthropic does let you switch on retention per workspace, so you can run Fable in a separate workspace while everything else stays on zero data retention. A request from a workspace that doesn't meet the requirement returns an error.
Where the data lives is a second question. The data residency page lists two options on the Claude API for where the model runs: global and us. There's no European option, and the only available location for stored data is the US. Google Cloud offers an EU multi-region endpoint at a 10 percent premium, but which Claude models it serves varies by region, so check Google's Model Garden before you assume Fable 5.2 is there. On Amazon Bedrock, Fable 5.2 runs only on the global and US endpoints.
For personal data, these are exactly the facts that belong in a DPIA: retention period at the processor, location, transfer outside the EU. Our DPIA step-by-step for an AI project shows where they go, and what EU data residency with a US provider does and doesn't settle is in EU data residency with US AI providers.
What breaks when you switch from Fable 5.1?
Moving from Fable 5.1 to 5.2 takes more than a new model name. Anthropic lists four changes that can break existing code.
- Editing earlier turns returns an error. Fable 5.2 checks that everything before a thinking block is unchanged: system prompt, tools and earlier messages. Code that trims history, rewrites it or swaps tools mid-conversation gets a 400 error with the message "The block is bound to a different conversation".
- Thinking blocks are tied to the model. Only Fable 5.2 and Mythos 5.2 read Fable 5.2's thinking blocks. A router that moves conversations between models loses the reasoning along the way, including when it moves to Opus 5.5.
- The old computer use tool no longer works.
computer_20251124returns an error everywhere except Amazon Bedrock. You move to the new toolset. - The advisor tool needs an equal model. A Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor.
Two rules apply to every request on top of that. Forced tool use (tool_choice set to a specific tool) isn't supported, and you can't switch adaptive thinking off. If you're coming from Fable 5.1, you already had those rules. From an older Opus model, you didn't.
The first point hits most custom integrations. Agents that manage their own context, summarise old messages to save tokens or adjust the system prompt per step do exactly what Fable 5.2 refuses. How to roll out that kind of change safely is in running AI agents in production.
Who is Claude Fable 5.2 right for?
Claude Fable 5.2 is right for a company already on Fable 5.1, or for work where Opus 5.5 demonstrably falls short. For nearly everything else, it's the wrong pick.
| Your situation | Advice |
|---|---|
| You run Fable 5.1 | Switch, after testing for the four changes above |
| You run Opus 5.5 and your test set passes | Stay; Fable costs double per token |
| You run Opus 5.5 and your test set fails on heavy reasoning | Test Fable 5.2, starting at high |
| You need zero data retention | No Fable; stay on Opus or Sonnet |
| You process personal data and require EU processing | Not through the Claude API; look at the EU route on Google Cloud or a model you host yourself |
Google made its own top model widely available a day later, at a temporarily low price; we cover it in Gemini 4 Carbon for business. The broader question of which model and which partner fit your organisation is in our guide to hiring AI consulting. If you want to know whether your integration survives the switch, or which model does your process cheapest per finished task, we work that out as part of AI consulting.