# Claude Fable 5.2 in Foundry: $8/$40 and the 30-day catch

Claude Fable 5.2 reached Microsoft Foundry on 9 November 2026 at $8 per million input tokens and $40 output, a 20% cut from Fable 5.1, with a separate Fast deployment at double the price. It is Anthropic-hosted only and a Covered Model that requires 30-day data retention, so ZDR organisations are locked out unless Anthropic authorises them, and the cost math against Opus 5.5 on Azure hosting decides whether the premium is worth it.

- Published: 2026-11-11 · Category: Azure & Cloud · Tags: Claude Fable 5.2, Microsoft Foundry, Anthropic, Azure AI, Claude Opus 5.5, Data Retention, GDPR, AI Agents, Prompt Caching
- Author: Technspire AB, Stockholm (https://technspire.com)
- Canonical: https://technspire.com/en/blog/claude-fable-5-2-foundry-8-40-and-the-30-day-catch

On 9 November 2026 Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, and Microsoft listed both in the Foundry model catalog the same day. The price moved this time: $8 per million input tokens and $40 per million output tokens, down from $10 and $50 on Fable 5.1. Cache reads fell from $0.25 to $0.20. A separate Fast deployment runs the same model at up to 2.5 times the output speed for exactly double the price. For an Azure team the catch sits in the fine print rather than the price sheet. Fable 5.2 has no Azure-hosted version; it runs on Anthropic infrastructure only. It is also a Covered Model that requires 30-day data retention. If your organisation negotiated zero data retention with Anthropic, requests to this model fail unless Anthropic has expressly authorised you. That one line decides whether the 20% saving is available to you at all.

## What shipped

Fable 5.2 is the general-release model, API ID `claude-fable-5-2`. Mythos 5.2 shares its weights, specification and pricing but ships with reduced cyber and biology safeguards, and only organisations verified through Anthropic's programmes such as the Cyber Verification Program can deploy it. On Foundry, Microsoft asks you to contact your account team to confirm Mythos eligibility. Both models keep the 1M token context window, the 128K maximum output and the tokenizer introduced with Opus 4.7. Adaptive thinking is always on, the default effort is `high`, and the knowledge cutoff is June 2026. Anthropic has committed not to retire the model before 9 November 2027. Fable 5.1 stays available and no deprecation has been announced for it.

Anthropic's own headline numbers are 30% faster output than Fable 5.1, roughly 55% fewer false-positive safeguard blocks, and 49.1% on FrontierCode 2.0 against 45.5% for Opus 5.5. The benchmark gap to Opus is narrower than the price gap, which matters for the decision guide below. Anthropic's model docs are explicit on this point: start on the latest Opus model, and move to Fable 5.2 only when your evals on Opus at a higher effort level still fall short.

### Price sheet, November 2026

| Model | Input / MTok | 5m cache write | Cache read | Output / MTok | Foundry hosting |
| --- | --- | --- | --- | --- | --- |
| Claude Fable 5.2 | $8.00 | $10.00 | $0.20 | $40.00 | Anthropic only |
| Claude Fable 5.2 Fast | $16.00 | $20.00 | $0.40 | $80.00 | Anthropic only |
| Claude Fable 5.1 | $10.00 | $12.50 | $0.25 | $50.00 | Anthropic only |
| Claude Opus 5.5 | $4.00 | $5.00 | $0.20 | $20.00 | Azure or Anthropic |
| Claude Sonnet 5.5 | $2.00 | $2.50 | $0.10 | $10.00 | Azure or Anthropic |

Two details in that table are easy to miss. Fable 5.2 and Opus 5.5 now charge the same $0.20 for a cache read, because Fable's multiplier is 2.5% of a higher base while Opus 5.5 uses 5% of a lower one. And the Fast row is not a flag you set on a request, as it is on the Claude API. On Foundry, Fast is a separate deployment named `claude-fable-5-2-fast`, so your routing layer picks the speed by choosing the deployment name. Batch processing is not available for the Fast deployment.

## The cost math on a long agent session

The saving is simplest to see on the workload Fable is built for: a tool-using loop that re-reads a growing prefix on every turn. Reuse the session model from [our Fable 5.1 post](/en/blog/claude-fable-5-1-foundry-cache-math-eu-caveats): 60 turns, a cached prefix averaging 150K tokens, 3K tokens of new tool results written to cache per turn and 2K tokens of output per turn. That is 9M cache-read tokens, 180K cache-write tokens and 120K output tokens.

| 60-turn session | Cache reads (9M tok) | Cache writes (180K tok) | Output (120K tok) | Total | CCU |
| --- | --- | --- | --- | --- | --- |
| Fable 5.1 | $2.25 | $2.25 | $6.00 | $10.50 | 1,050 |
| Fable 5.2 | $1.80 | $1.80 | $4.80 | $8.40 | 840 |
| Fable 5.2 Fast | $3.60 | $3.60 | $9.60 | $16.80 | 1,680 |
| Opus 5.5 | $1.80 | $0.90 | $2.40 | $5.10 | 510 |
| Sonnet 5.5 | $0.90 | $0.45 | $1.20 | $2.55 | 255 |

The Fable 5.1 to 5.2 move is a flat 20% across every line, since all three rates fell by the same factor. The more useful comparison is the Opus 5.5 row. On the price sheet Fable 5.2 costs double. On this cache-heavy session it costs 65% more, because the cache reads that dominate a long loop are priced identically. That premium buys the FrontierCode gap and the longer autonomous runs Anthropic describes. Whether it is worth 330 extra CCUs per session is a question your own evals answer, not the benchmark table.

The Fast row is the one to be disciplined about. Cache multipliers stack on top of fast-mode pricing, so every line doubles and the session lands at $16.80 for the same tokens. You get output at up to 2.5 times the speed. That is a good trade for an interactive coding session where a developer is waiting on each turn. It is a poor trade for an overnight batch of document analyses, which nobody is watching and which would otherwise qualify for the batch rate of $4 and $20 that Microsoft lists for the standard deployment.

On Foundry all of this bills in Claude Consumption Units at a fixed $0.01 each, metered hourly to the Azure Marketplace and invoiced monthly in arrears. Anthropic rates your tokens at list price, applies any negotiated discount, then converts. The CCU column above is what your Azure Cost Management view shows for each session. Nothing else changed in the billing model since Fable 5.1, so the MACC and quota notes in that earlier post still apply, and the [Opus 5.5 migration post](/en/blog/claude-opus-5-5-foundry-opus-5-migration-cost-math) covers the Azure-hosted side of the comparison.

## What Foundry actually gives you

Foundry offers Claude in two hosting versions. Hosted on Azure runs inference on Azure infrastructure end to end, and Anthropic's own documentation states that for those deployments prompts and completions remain within Azure, with only usage metadata and safety-flagged content egressing to Anthropic. Hosted on Anthropic routes the request to Anthropic's infrastructure, with Microsoft providing the portal and the invoice. The Azure-hosted option exists for Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5, Sonnet 5, Haiku 5.5 and Haiku 4.5. Every Fable and Mythos model, 5.2 included, is Anthropic-hosted only.

That choice has consequences beyond data flow. Anthropic-hosted deployments get the richer feature set: code execution, Agent Skills, the Files API, programmatic tool calling and the current web search and web fetch tool versions. Azure-hosted deployments return a 400 for all of those by design. So a team that wants Fable 5.2 for autonomous research agents gets the tools it needs, while a team that chose Azure hosting for Opus 5.5 to keep prompts inside Azure gives some of them up. Several things are missing on Foundry whichever hosting you pick: the Message Batches API, server-side refusal fallback, the advisor tool, the Models API and the Admin and Compliance APIs.

Two of Microsoft's statements do not line up with Anthropic's. Microsoft's launch post lists a batch rate for Fable 5.2 and names both Global Standard and US DataZone as deployment options. Anthropic's Foundry page says the Batches API is unsupported on Foundry and that Anthropic-hosted deployments are Global Standard only, with US Data Zone reserved for Azure-hosted models. Treat the batch rate and the DataZone option as unconfirmed until you see them in your own portal. If a US Data Zone deployment does appear, expect the 1.1 times multiplier that Anthropic applies to US-only inference everywhere else.

### Calling it

The endpoint shape is unchanged: your Foundry resource name, the Anthropic path prefix and the deployment name in the model field. Entra ID is the sensible default for anything beyond a prototype. Effort stays the main cost lever, and the API default of `high` is where Anthropic says most of the capability already sits.

```
import AnthropicFoundry from "@anthropic-ai/foundry-sdk";
import { DefaultAzureCredential, getBearerTokenProvider } from "@azure/identity";

const client = new AnthropicFoundry({
  resource: "contoso-foundry",
  azureADTokenProvider: getBearerTokenProvider(
    new DefaultAzureCredential(), "https://ai.azure.com/.default"),
});

const response = await client.messages.create({
  model: "claude-fable-5-2",          // or "claude-fable-5-2-fast"
  max_tokens: 64000,
  output_config: { effort: "high" },  // low | medium | high | xhigh | max
  messages: [{ role: "user", content: "..." }],
});

if (response.stop_reason === "refusal") {
  // safeguards fired: retry on claude-opus-5-5 client-side
}
```

The refusal branch is not optional. Fable 5.2 runs cyber and biology classifiers, and a declined request returns HTTP 200 with a refusal stop reason and a category. On the Claude API you can ask the server to retry on Opus 4.8 or Opus 5 automatically. Foundry does not support that parameter, so the retry is your code's job, and the SDKs ship a client-side fallback middleware for exactly this. The good news in this release is the volume: Anthropic reports roughly 60% fewer false-positive cyber blocks than on 5.1, which is the difference between a fallback path that runs daily and one that runs weekly.

## The 30-day catch

Anthropic designates Fable 5.2 and Mythos 5.2 as Covered Models, alongside the 5 and 5.1 generations. Covered Models require 30-day data retention for safety monitoring, and zero data retention is not available for them unless Anthropic expressly authorises it. The requirement applies wherever the models are offered, Foundry included. Microsoft's launch post phrases the same rule from the other direction: until Enterprise Frontier Safeguards is available, eligible customers can use Fable 5.2 with zero data retention. Eligible means authorised by Anthropic. Everyone else accepts 30 days.

Who holds those 30 days is the part to get in writing. Anthropic acts as an independent processor for Microsoft, and customers using Claude through Foundry are subject to Anthropic's data use terms. On the Anthropic-hosted path, inference runs on Anthropic's infrastructure, so the retained prompts and completions sit with Anthropic, not in your Azure subscription. Anthropic's retention page confirms this for the Claude API and the Anthropic-processor platforms, and contrasts it with Bedrock and Google-managed deployments, where retained data stays inside the cloud provider's environment. There is no Google-managed equivalent on Foundry for Fable 5.2, because there is no Azure-hosted version.

If your organisation has a ZDR arrangement with Anthropic and still wants Fable 5.2, Anthropic's documented route on the Claude API is to enable standard 30-day retention for one named workspace, leaving the rest of the organisation on ZDR. Ask your Anthropic and Microsoft account teams how that override maps to a Foundry resource before you plan around it, because the Foundry documentation does not describe the equivalent switch. Enterprise Frontier Safeguards, which Microsoft references as the eventual fix, is the arrangement where safety monitoring runs against data held in the customer's own environment. It is not generally available today and has no published date.

## Migrating from Fable 5.1: what breaks

Anthropic lists four breaking changes, and three of them can reach a Foundry integration.

**The prefix check now applies to every account.** Fable 5.2 validates that nothing before a replayed thinking block has changed since the model produced it: not the system prompt, not the tool list, not an earlier message. On Fable 5.1 that check only fired for accounts created on or after 31 August 2026, so older accounts could edit history freely and never see the error. On 5.2 the same code returns a 400 whose message says the block is bound to a different conversation. If your harness trims old tool results, rewrites earlier turns or swaps the system prompt mid-conversation, test it before you switch. The stopgap is the `thinking-binding-controls-2026-08-01` beta with `prefix_mismatch_behavior` set to drop the block, which costs you the dropped reasoning and can raise output tokens.

**Only 5.2 models read 5.2 thinking blocks.** On the Claude API, Anthropic documents that Fable 5.2 can pick up a conversation that started on Opus 5.5 or Fable 5.1 and keep their reasoning. The reverse is not true on any platform. Move a Fable 5.2 conversation to Opus 5.5, to Fable 5.1 or to the Opus 4.8 and Opus 5 fallback models, and every Fable 5.2 thinking block is dropped before the model sees it. The request still succeeds and the dropped tokens are not billed, but a router that bounces between models for cost reasons silently loses the reasoning it paid for. Design the fallback as a one-way door.

**The old computer use tool is gone.** A request that declares `computer_20251124` returns a 400 on every platform except Bedrock. Declare the `computer_toolset_20260801` toolset instead. The fourth change concerns the advisor tool, which Foundry does not support, so it cannot affect you there.

Two behavioural shifts need prompt work rather than code. Anthropic's testing found that instructions asking the model to answer without deliberating no longer stop it thinking at medium effort and above, so a prompt line is not a cost control any more; lower the effort for those turns instead. And at low and medium effort the model more often ends a long task with a list of what remains rather than finishing it. If an agent must run to completion, say so in the prompt and keep effort at high.

## Fable 5.2 or Opus 5.5: the decision

| Question | Opus 5.5 on Azure hosting | Fable 5.2 on Anthropic hosting |
| --- | --- | --- |
| Where do prompts go? | Stay within Azure; metadata and flagged content to Anthropic | Anthropic infrastructure; Anthropic is the processor |
| Can I keep inference in one geography? | US Data Zone at 1.1x; no EU zone | Global Standard per Anthropic; no EU zone |
| Zero data retention? | Per your Anthropic arrangement | No, unless Anthropic authorises it |
| Code execution, Skills, Files API? | Not on Azure hosting | Yes |
| Fast deployment? | Yes, on Azure hosting | Yes, separate deployment |
| 60-turn session above | $5.10 (510 CCU) | $8.40 (840 CCU) |

Opus 5.5 on Azure hosting is the default for a European enterprise: it is the only way to keep prompts inside Azure, it costs 39% less on a cache-heavy session, and it leaves your retention arrangement untouched. Fable 5.2 earns its place when a specific agent fails your evals on Opus 5.5 at xhigh effort, or when it needs code execution and Skills that Azure hosting does not expose. In that case, run it from a separate Foundry resource and a separate workspace, so the retention and processor terms attach to one agent rather than your whole estate.

## The Swedish and EU angle

Nothing about Fable 5.2 changes the geography of Claude on Azure. No Claude model in Foundry has an EU data zone, and that was true before this release. Creating the Foundry resource in Sweden Central gives you a Swedish endpoint and a place on your MACC, but a Global Standard deployment can run inference anywhere. For the Anthropic-hosted path the request leaves Azure altogether. Your records of processing need to say so, and your Article 28 chain now has Microsoft and Anthropic in it, with Anthropic holding 30 days of prompts under its own terms.

For public-sector and regulated buyers that is the same analysis [Regeringskansliet made in October](/en/blog/regeringskansliet-blocked-chatgpt-claude-kept-copilot-azure-playbook) when it blocked direct ChatGPT and Claude access but kept Microsoft 365 Copilot. The deciding factor there was not the model but who processes the data and under which contract. Fable 5.2 through Foundry lands on the Anthropic side of that line, with a Microsoft invoice. If your DPIA for Opus 5.5 relied on the Azure-hosted promise that prompts and completions stay within Azure, do not copy it across to Fable 5.2. It needs its own assessment, a transfer analysis for the US processing, and a note on the retention period and purpose, which Anthropic states is safety monitoring.

On the AI Act side the obligations are unchanged by the version bump. Anthropic carries the general-purpose model duties. You carry deployer duties for the system you build, and the Article 50 transparency rules have applied since 2 August 2026.

## Decision guide

- **Check your retention arrangement first.** If you have ZDR with Anthropic, Fable 5.2 will fail until you have either an authorisation from Anthropic or a workspace with 30-day retention enabled. Settle this before anyone deploys.
- **Deploy both names.** Create `claude-fable-5-2` and `claude-fable-5-2-fast` as separate deployments and route by latency need. Keep Fast off batch and background jobs.
- **Run the prefix check test.** Replay a week of real conversations from your harness against the new deployment before cutover. A 400 citing a bound block means your code edits history somewhere.
- **Make fallback one-way.** Route refusals to Opus 5.5 client-side and accept that the Fable reasoning is lost for those turns. Do not round-trip conversations back to Fable 5.2 expecting the context to survive.
- **Re-run evals at high, then xhigh.** Anthropic's advice is to start at the default and step up only where you measure a gain. The 65% session premium over Opus 5.5 must show up in your pass rate.
- **Verify the Microsoft claims in your portal.** The batch rate and the US DataZone option appear in Microsoft's post but not in Anthropic's Foundry documentation. Plan on neither until you see them.
- **Isolate the deployment.** Separate Foundry resource and a separate DPIA. The processor and retention terms are different from your Azure-hosted Claude models and should be documented that way.

## Sources

- [Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2](https://www.anthropic.com/claude-fable-and-mythos-5-2)
- [Microsoft Foundry blog: Claude Fable 5.2 and Mythos 5.2 bring frontier intelligence to Microsoft Foundry](https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/claude-fable-5-2-and-mythos-5-2-bring-frontier-intelligence-to-microsoft-foundry/4564282)
- [Claude Platform docs: Claude Fable 5.2 overview](https://platform.claude.com/docs/en/models/fable-5-2/overview)
- [Claude Platform docs: What's new in Claude Fable 5.2](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2)
- [Claude Platform docs: Claude in Microsoft Foundry](https://platform.claude.com/docs/en/build-with-claude/claude-in-microsoft-foundry)
- [Claude Platform docs: Pricing, including Foundry CCU billing and fast mode](https://platform.claude.com/docs/en/about-claude/pricing)
- [Claude Platform docs: API and data retention, model-specific requirements](https://platform.claude.com/docs/en/manage-claude/api-and-data-retention)
- [Claude Platform release notes, 9 November 2026](https://platform.claude.com/docs/en/release-notes/overview)

---

Technspire AB builds AI agents, Azure OpenAI solutions, and production web platforms for Swedish and EU enterprises. Book a call: https://calendly.com/technspire · hello@technspire.com · More articles: https://technspire.com/en/blog · Site overview for agents: https://technspire.com/llms.txt