Claude Fable 5.5 Landed on Azure First: How to Read a Catalogue Row With Nothing Behind It

On October 13, 2026, a model that has never been announced acquired a lifecycle status. Azure’s models API in eastus2 began returning Claude Fable 5.5 as A​nthropic version 1, with a lifecycleStatus of GenerallyAvailable and a deprecation date of December 5, 2027. The same day, LiteLLM’s provider-sync bot opened and merged pull request #46201, adding exactly one row for azure_ai/claude-fable-5-5 to its price-and-context map. Nothing else moved. There is still no entry on A​nthropic’s models overview, no newsroom post, no benchmark run, and no price published anywhere by the lab that would have to publish one. Claude Fable 5.1 — the model the new row is filed under, and the one you can actually call today — remains the newest shipped member of the family. So the interesting question is not “is this real.” It is what a cloud catalogue row actually proves, and what a dozen people reading it this week will wrongly conclude it proves.

What changed on October 13

Two things, and only two. First, the Azure management API endpoint that lists models for a region — locations/eastus2/models at API version 2024-10-01 — started including a claude-fable-5-5 entry that was not there before. Its lifecycleStatus is GenerallyAvailable, not Preview. That distinction matters because the two Fable-family models Azure already carries, Claude Fable 5 and Claude Fable 5.1, are both still filed as Preview in the same catalogue. A GA flag is Azure saying the service is supported and billable under its normal terms, not Microsoft saying the underlying model has launched.

Second, LiteLLM mirrored the listing. Its sync bot generates pull requests from provider model APIs rather than from hand-written tables, and this one adds a single JSON object. The PR body is unusually candid about the limits of its own evidence, and that honesty is the most useful thing in it: the bot states that no source states a price, context window or capability for Fable 5.5 yet, that the Azure Claude models page, the Azure models-sold-directly page, A​nthropic’s pricing page and models overview, and the Bedrock, Vertex and other catalogues it checks all have no Fable 5.5 entry, and that the pricing API returned no structured row for it. Every field in the row it just added is marked inherited_from: azure_ai/claude-fable-5-1.

Which is a strange but coherent thing for a cost map to do. LiteLLM is a budget-tracking layer; a provider catalogue entry with no price is a crash waiting to happen, so the tool copies the closest known sibling. It is not evidence that Fable 5.5 costs what Fable 5.1 costs. It is evidence that someone needed a placeholder.

The row itself, value by value

Here is what the merged diff actually contains, and where each figure comes from. Read the second half of every line — that is the part people will skip.

• Price — the row carries $10 per million input tokens and $50 per million output tokens, and every cost field is marked inherited from azure_ai/claude-fable-5-1, whose published rate card is $10 and $50 per million.

• Context window — 1,000,000 max input tokens and 128,000 max output tokens, both inherited rather than sourced.

• Prompt caching — $0.25 per million cached reads, $1.25 per million cache writes, $2.00 per million for the one-hour cache, with a 512-token minimum cacheable prompt. All four inherited.

• Thinking — thinking_always_on: true and supports_adaptive_thinking: true, meaning the model would reason by default rather than on request. Inherited, and a notable inheritance: it is the single field most likely to change between a 5.1 and a 5.5.

• Modalities and tools — vision, PDF input, function calling, tool choice, native structured output, response schemas and mid-conversation system prompts are all true; assistant prefill, forced tool use and general sampling parameters are all false. Inherited.

• Provenance — the one field the bot did not inherit is the source URL, which points at the Azure management endpoint for eastus2.

Two details deserve more weight than they are getting. The deprecation date, December 5, 2027, is roughly three weeks later than the retirement floor A​nthropic has published for its current generation, which is what you would expect if this is one more model in the same generation rather than a new lineage. And the GenerallyAvailable status is region-scoped: this is an eastus2 observation, not a global one. A model can sit in one region’s catalogue and nowhere else for a long time.

Where Claude Fable 5.5 still does not exist

Worth stating plainly, because the listing is easy to over-read. As of October 13, 2026, checking the surfaces a shipped model has to appear on:

• A​nthropic’s newsroom — the most recent post is for Claude Haiku 5.5 on October 7, 2026. There is no Fable 5.5 announcement.

• A​nthropic’s models overview — the current lineup is listed as Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5. Fable 5.5 is absent, and the documentation sitemap has no fable-5-5 path.

• Azure’s own customer-facing documentation — the models-sold-directly page does not mention Fable at all, in any version. Microsoft’s AI Foundry catalogue listing for A​nthropic still shows only Claude Fable 5 and Claude Fable 5.1, both marked Preview. A management-API entry for a model that the sales documentation has never heard of is a real gap, and it cuts both ways: it means the listing is early, and it means the listing may not survive contact with the product team.

• Independent measurement — Artificial Analysis has no page for Fable 5.5, and no scores, which is expected and tells us nothing except that nobody outside has run it.

• Our own catalogue — OrcaRouter’s model API lists 207 models and does not include anthropic/claude-fable-5.5. It does include anthropic/claude-fable-5 and anthropic/claude-fable-5.1. We are not going to pretend otherwise: there is nothing to serve, so we do not list it.

A generally-available listing is not a release

The confusion to avoid is treating GenerallyAvailable as a launch. It is a support and billing classification inside one cloud. The thing that would actually constitute a release — a model card, a rate card, a context window A​nthropic will stand behind, a model identifier you can pass to an endpoint — is absent on every one of those five checks.

What the listing is good evidence for is narrower and more interesting: that an A​nthropic successor to Fable 5.1 has cleared Microsoft’s internal readiness process far enough to get a lifecycle status in at least one region, and that Microsoft’s tooling exposes it to anything reading the management API. That is a meaningful step. It is just not the same step as “available.”

The signal post that surfaced this — an X post from @kimmonismus on the afternoon of October 13 — also folds in a story about A​nthropic’s IPO and an expectation that the release will follow other labs’ recent models. Treat that part as speculation by the poster. Nothing in the Azure entry or the LiteLLM diff speaks to capital markets, and the ordering claim is a guess. The API string is the evidence; the narrative wrapped around it is not.

What you can verify yourself today

The useful response to a catalogue sighting is not to argue about it. It is to know which checks settle the question, and to run them again at intervals rather than once.

• Query the Azure management endpoint for your region and grep the returned list for the string. A row that appears, disappears and reappears over a week tells you the entry is unstable; a row that persists for a fortnight is a different proposition.

• Try the identifier against an endpoint. If a provider accepts claude-fable-5-5 and returns a response you can read a model field from, the situation has changed materially and this page is out of date. Until then, a catalogue entry with no route is just metadata.

• Watch the vendor’s own documentation sitemap rather than its marketing pages. Documentation paths tend to appear hours before a newsroom post, and they are machine-readable.

• Re-check the flag. If GenerallyAvailable reverts to Preview or the entry vanishes, that is the strongest possible signal that the listing ran ahead of the product, and it is worth knowing before you build a plan on it.

• Refuse to inherit in your own head. The single most likely error a reader makes this week is quoting $10 and $50 per million as “Fable 5.5 pricing.” That number belongs to Fable 5.1 until A​nthropic publishes otherwise.

If it does ship, what changes on your bill

There is a version of this story worth planning for, and it does not require betting anything. If Fable 5.5 does launch, the questions that will decide whether you adopt it are the ones nobody can answer yet: whether reasoning is always on, what it costs relative to Fable 5.1, and whether its speed holds up in production. Two of those three are the exact fields that were inherited rather than sourced, which is a decent argument for not pricing a roadmap against this row.

What is true today, and is a reason the decision is low-stakes, is that the shipped member of the family is already reachable behind a single OpenAI-compatible endpoint alongside 207 other models on OrcaRouter, with no second contract and no code change when you want to compare them. We pass through provider list pricing at a 0% markup, which means that when a vendor cuts a price the change is live on our side the same day rather than whenever a table gets rebuilt. And for a model that has cleared a cloud catalogue but has no published behaviour, automatic failover is the difference between “I tried it and it broke my afternoon” and “I pointed a fraction of traffic at it and it fell back.”

The honest summary: the strongest evidence yet that a Fable 5.5 exists now sits in an Azure region list and a copied JSON object, and the strongest evidence that it is not shipping yet sits in the fact that the company that makes it has not written a single sentence about it. Both of those are true on the same day, and the second one is the one that should govern what you do this week. What to watch is not the newsroom. It is whether the eastus2 row is still there on October 27, whether its status still says GenerallyAvailable, and whether anyone publishes a number that someone other than a placeholder bot will stand behind.

Whether a catalogue row proves a model exists is exactly the kind of question the documented AI failures and agent incidents in the OrcaRouter AI Incident Archive were collected to answer, where each entry carries its own sources, severity, confidence grade and disputed flag.