Claude Fable 5.1 in Foundry: cache math and the EU caveats

On 1 September 2026 Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1. The headline price did not move: $10 per million input tokens and $50 per million output tokens, the same as Fable 5. The number that did move is the cache read rate, which dropped from $1 to $0.25 per million tokens. The same day, Microsoft listed Fable 5.1 in the Foundry model catalog and started rolling it into Copilot Cowork and Copilot Studio. If you run agents on Azure from Sweden, three questions matter more than the benchmark table: what a long agent session now costs, where your prompts are processed, and what breaks if you are already on Fable 5. The answers are less tidy than the launch posts suggest.

What shipped

Fable 5.1 is the general-release model, API ID claude-fable-5-1. Mythos 5.1 shares the same weights, specs and pricing but runs with reduced cyber and biology safeguards and is offered only to approved organisations in Anthropic's Project Glasswing and trusted access programmes. Both carry a 1M token context window at flat per-token pricing, 128K max output, adaptive thinking that is always on, and the tokenizer introduced with Opus 4.7. The reliable knowledge cutoff is June 2026. Anthropic commits not to retire Fable 5.1 before 1 September 2027 on its own platforms, which include Foundry.

Anthropic's own estimate is that Fable 5.1 costs about 25% less than Fable 5 for typical workloads and up to about 45% less for heavily agentic ones. Every cent of that saving comes from cache reads, so the claim only holds if your prompts actually hit the cache. The capability gains concentrate in long agentic coding sessions, document and spreadsheet work, multistep research, dense-chart vision and full-window long-context reasoning. On Terminal-Bench 4.0 Anthropic reports 55.8% for Fable 5.1 against 42.0% for Fable 5 and 52.3% for Opus 5. Treat vendor benchmarks as a direction of travel, not a procurement input.

Price sheet, September 2026

Model	Input / MTok	5m cache write	Cache read	Output / MTok
Claude Fable 5.1	$10.00	$12.50	$0.25	$50.00
Claude Fable 5	$10.00	$12.50	$1.00	$50.00
Claude Opus 5	$5.00	$6.25	$0.50	$25.00
Claude Sonnet 5	$2.00	$2.50	$0.20	$10.00

Cache reads on Fable 5.1 are priced at 0.025 times base input, where every other Claude model uses 0.1. That makes a Fable 5.1 cache hit cheaper in absolute terms than an Opus 5 cache hit, even though Fable's base price is double. One-hour cache writes stay at $20 per million tokens and the 512-token minimum cacheable prefix is unchanged. Sonnet 5 keeps its $2/$10 introductory price permanently; the September increase has been cancelled.

The cache math for a real agent session

The saving is easiest to see on the workload Fable is sold for: a long tool-using loop where the system prompt, tool definitions and growing history are re-read on every turn. Take a 60-turn coding or research session. Assume the cached prefix averages 150K tokens per turn (it starts small and grows), each turn adds 3K tokens of new tool results that get written to the 5-minute cache, and the model emits 2K output tokens per turn including its thinking. Ignore the initial write and any cache misses from gaps longer than five minutes.

60-turn session	Cache reads (9M tok)	Cache writes (180K tok)	Output (120K tok)	Total
Fable 5	$9.00	$2.25	$6.00	$17.25
Fable 5.1	$2.25	$2.25	$6.00	$10.50
Opus 5	$4.50	$1.13	$3.00	$8.63
Sonnet 5	$1.80	$0.45	$1.20	$3.45

Two things fall out of this. The Fable 5 to Fable 5.1 move cuts the session by 39%, which sits inside Anthropic's 25 to 45% range. More interesting is the gap to Opus 5. On the price sheet Fable costs exactly double. On this cache-heavy session it costs 22% more, because the cache reads that dominate a long loop are now cheaper on Fable than on Opus. The premium you pay for Fable 5.1 shrinks as sessions get longer and re-read more context. Short, chatty requests with little cached prefix still pay the full 2x.

On Foundry that $10.50 session shows up on your Azure invoice as 1,050 Claude Consumption Units, since one CCU is a fixed $0.01 and Anthropic rates your tokens at the list prices above before converting. CCU spend counts against a Microsoft Azure Consumption Commitment, which is usually why Claude goes through Foundry rather than Anthropic's own API. For the mechanics of cache writes, refreshes and the 1-hour tier, see our earlier breakdown of Anthropic prompt caching pricing.

What Foundry actually gives you

Foundry offers Claude in two hosting versions. Version 2, "Hosted on Azure", runs inference on Azure infrastructure end to end and is what you get for Opus 5, Opus 4.8, Sonnet 5 and Haiku 4.5. Version 1, "Hosted on Anthropic infrastructure", routes your request to Anthropic's own servers with Microsoft providing the portal, the endpoint and the bill. Fable 5.1 is version 1 only. There is no Azure-hosted Fable, and Microsoft's model page still marks it as preview even though the Foundry launch blog calls it generally available. Check the catalog badge before you promise anyone a production SLA.

Deployment is Global Standard only. The region table lists Fable 5.1 in exactly two Azure regions, East US 2 and Sweden Central, so a Stockholm team can create the Foundry resource next to its other workloads. Do not read that as data residency. Microsoft's own wording for Anthropic-hosted models is that data "might be processed outside Azure, including outside the selected Azure region." The Sweden Central label governs where your Foundry resource and its metadata live, not where Fable 5.1 runs.

Quota: pay-as-you-go starts at zero

The default rate limits table is the detail most teams will trip on. For Fable 5.1 and Fable 5, pay-as-you-go subscriptions get 0 requests per minute, 0 input tokens per minute and 0 output tokens per minute by default. You must file the Foundry quota increase request before a single call succeeds. Enterprise Agreement and MCA-E subscriptions get 4,000 RPM, 4 million uncached input tokens per minute and 800,000 output tokens per minute. Cache reads do not count toward the input limit, so a session that is mostly cached prefix burns almost none of your ITPM.

Features that do not exist on Foundry

No Message Batches API. The 50% batch discount that takes Fable 5.1 to $5/$25 is a Claude API and Bedrock feature. On Foundry every token is billed at the synchronous rate.
No server-side fallback. The fallbacks parameter that reroutes a refused request to Opus 5 or Opus 4.8 inside one call is unavailable. Use the client-side fallback middleware in the Anthropic SDKs, or write your own retry on stop_reason: "refusal".
No Models API, Admin API, Compliance API or Managed Agents. Use Azure Monitor, Log Analytics and Cost Management for the operational side, and note that Anthropic's rate-limit headers are not returned either.
Fixed sampling. temperature and top_k are rejected, top_p must be at least 0.99, and thinking may only be adaptive or omitted. If your Azure OpenAI code sets temperature: 0 for determinism, it will 400 against Fable.

Authentication is either an Azure API key or an Entra ID bearer token against https://<resource>.services.ai.azure.com/anthropic/v1/messages. The C#, Java, PHP, Python and TypeScript SDKs have native Foundry clients. Go and Ruby do not, and need a base URL override.

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

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

const res = await client.messages.create({
  model: "claude-fable-5-1",          // deployment name
  max_tokens: 16000,
  output_config: { effort: "high" },  // low | medium | high | xhigh
  cache_control: { type: "ephemeral" },
  system: STABLE_SYSTEM_PROMPT,
  tools: TOOLS,                       // tool_choice stays "auto"
  messages,
});

if (res.stop_reason === "refusal") {
  // Foundry has no server-side fallbacks; reroute here
  console.warn(res.stop_details?.category);
}

Note that the max effort level is not offered for Fable 5.1 on Foundry; xhigh is the ceiling. Microsoft's table also lists the per-turn effort, mid-conversation system message and token budget betas for Opus 5 and Sonnet 5 but not for Fable 5.1, so test those before a design depends on them. For the base deployment walkthrough and how CCU metering appears in Cost Management, see Claude on Azure is GA.

Data residency and retention: the two caveats

There is still no EU data zone for any Claude model

Foundry's Data Zone Standard deployment type, the one that keeps Azure OpenAI traffic inside EU member states, is available for Claude only as a US zone and only for the Azure-hosted Opus 5, Opus 4.8 and Sonnet 5. The Europe tab of the Data Zone table reads "Not available." Anthropic's regional compliance page lists Microsoft Foundry in Europe as "Coming 2026", while Bedrock and Vertex AI already have EU regional endpoints with processing guarantees. With four months left in the year, plan on the assumption that it slips.

For Fable 5.1 specifically the situation is one step further out. Being Anthropic-hosted, it cannot use even the US data zone, and Anthropic is the independent data processor under its own Data Processing Addendum rather than Microsoft's. If your DPIA for an Azure OpenAI workload leaned on "processed within the EU Data Zone", that paragraph does not transfer. You are documenting a transfer to a US processor with its standard contractual clauses, exactly as you would for Anthropic's first-party API.

30-day retention is mandatory, for now

Fable 5.1, Mythos 5.1, Fable 5 and Mythos 5 are what Anthropic calls Covered Models. They require 30-day retention of prompts and outputs for safety monitoring and are not available under zero data retention unless Anthropic expressly authorises it. Organisations with a ZDR contract can carve out a single workspace with 30-day retention for these models and keep ZDR elsewhere. That retention sits with Anthropic wherever Anthropic is the processor, which includes Foundry.

The same day as the model launch Anthropic announced Enterprise Frontier Safeguards, its planned replacement for that rule. Under EFS the activity data used for misuse monitoring is stored in the customer's own cloud account, including Azure Blob Storage, under the customer's encryption keys and access policies, and automated systems analyse a rolling window of traffic with no human review by Anthropic staff. Microsoft Foundry is on the supported platform list. Anthropic charges nothing for EFS; you pay your cloud provider for the storage. Rollout is phased, with broad availability targeted for "later this fall", and eligible customers get zero data retention on Fable 5 and 5.1 in the meantime. If retention is the blocker for your security team, ask your Anthropic or Microsoft account team about EFS eligibility now rather than waiting for general availability.

Safeguards, refusals and Mythos

Fable 5.1 runs input and output classifiers for dual-use cyber and biology content. When one fires you get HTTP 200 with stop_reason: "refusal" and a stop_details object naming the category, and you are not billed for the refused input tokens. Anthropic says the cyber classifier now intervenes about 60% less often per session than on Fable 5 and permits vulnerability discovery while still blocking exploit development, and the biology classifier fires 85% less often on elementary and medical questions. For a Swedish SOC or a medtech team that hit false positives on Fable 5, that is worth re-testing.

Mythos 5.1 is also listed in the Foundry catalog, as a gated research preview with Entra ID authentication only. Access is granted at Anthropic's discretion and prioritised for defensive cybersecurity. Anthropic's launch post routes access through its Cyber Verification Program and a Life Sciences Verification Program developed with the US government. For most EU organisations the realistic path is Fable 5.1 with its classifiers, plus a fallback model for the categories where they fire, as described in Locking down AI agents.

Migrating from Fable 5: three things break

Changing the deployment name is the easy part. Anthropic's migration notes list three breaking changes and several behaviour shifts.

Forced tool use returns 400. tool_choice of type any or tool is rejected, on token counting too. Keep auto, set strict: true on the tool for schema-valid arguments, or move the schema to structured outputs. Name the tool in the prompt when you need it called.
Thinking blocks are bound to the model that made them. Fable 5.1 can read Opus 5 or Fable 5 thinking blocks, but no earlier model can read Fable 5.1's. A router that falls back mid-conversation loses the reasoning for the turns it runs elsewhere. The drop is silent unless you send the thinking-binding-controls-2026-08-01 beta header.
Editing history invalidates thinking. Rebuilding the system prompt or tool list between requests, injecting and then deleting a per-turn reminder, or reordering earlier turns invalidates every later thinking block. Accounts created on or after 31 August 2026 get a 400 on replay. Older accounts are only warned unless they opt in. Make your harness append-only and use mid-conversation system messages for changes.

The behaviour shifts matter for cost. Fable 5.1 batches fewer independent tool calls per turn than Fable 5 did, which adds round trips unless you add a one-line batching instruction. It writes fewer progress notes between tool calls, so a UI that showed narration goes quiet unless you request display: "updates". And it is more likely to rewrite a whole file than make a targeted edit, at $50 per million output tokens. Re-run your evals and your cost dashboards on the same day.

The Copilot side

Microsoft's Copilot blog says Fable 5.1 began rolling out to eligible users in Copilot Cowork and Copilot Studio on 1 September, with availability varying by region and organisation, and administrators controlling access from the Microsoft 365 admin center. Nothing in the announcement changes the data-flow picture from when Sonnet 5 arrived in Copilot: Anthropic-model requests are processed by Anthropic under Anthropic's terms. EU tenants that documented that flow in July can reuse the assessment, and tenants that switched Anthropic models off will find Fable 5.1 off as well. Our EU tenant guide to Claude in Microsoft 365 Copilot walks through the settings.

The Swedish and EU angle

Article 50 marking is handled at the model. Microsoft's Foundry model page states that Fable 5.1 and Mythos 5.1 comply with the EU watermarking standard using a double-key scheme: an interwoven text watermark with a second key for EU detection, plus C2PA content credentials on generated media. Watermarking happens server-side at generation time with no change to request or response shapes. The detection API is in early access and outside Foundry's launch scope, so you can rely on the mark for your own transparency obligations but cannot yet verify it yourself. See our Article 50 analysis of Claude's text watermark for how that maps onto deployer duties.

Residency decides the platform, not the model. If a workload must process inside the EU, Fable 5.1 on Foundry is not a candidate today. The choices are an EU regional endpoint on Bedrock or Vertex AI, where Fable 5.1 is also available, or an Azure-hosted Opus 5 or Sonnet 5 deployment with "Global Standard" processing anywhere Azure runs. Swedish public-sector buyers who standardised on Foundry for MACC and Entra ID reasons should write the EU-zone gap into the risk register with a review date in Q1 2027, since "Coming 2026" is the only public commitment.

Retention is a contract question before it is a technical one. Thirty days of prompt retention at a US processor is defensible for many enterprise workloads and indefensible for some, such as patient data or classified procurement material. EFS moves that data into your own Azure Blob Storage under your keys, which is a materially better story for a DPIA, but it is not available yet and eligibility is decided by Anthropic. Anthropic says it designed EFS with more than 100 customers in financial services, healthcare, manufacturing, telecoms, law, retail and government, so a regulated Swedish enterprise is squarely the intended audience.

Decision guide

Already on Fable 5 via Foundry? Migrate. Same price, cheaper cache, better model, one-year retirement runway. Budget a sprint for the three breaking changes and re-tune effort from the default high.
On Opus 5 and considering the jump? Measure your cache hit ratio first. If long sessions re-read most of their context, Fable 5.1 costs roughly 20 to 25% more than Opus 5, not 100%. If your traffic is short and uncached, stay on Opus 5 at higher effort until your evals say otherwise.
Need EU processing? Not on Foundry, for any Claude model, in any version. Use Bedrock or Vertex EU endpoints, or Azure-hosted Opus 5 and Sonnet 5 with the US data zone if US-only processing is acceptable.
Blocked on retention? Ask about EFS eligibility now. In the transition, eligible customers get ZDR on Fable 5 and 5.1.
Pay-as-you-go subscription? File the quota request on day one. Your default Fable quota is zero.
Batch workloads? Foundry has no Batches API. Route offline jobs through the Claude API or Bedrock for the 50% discount, or accept synchronous pricing for the MACC benefit.