Claude Fable 5.5 Leaks: Azure Row, Onigiri, Checkpoints

Today, October 20, 2026, is the Tuesday the Claude Fable 5.5 rumor picked for itself. Last week several X accounts said “next Tuesday,” and every Fable launch so far has landed on a Tuesday. As of this morning, though, the shipped Fable model is still Claude Fable 5.1. What turned up overnight was a sentence from one X account saying a “second checkpoint” went to partners on Sunday.

This post is for developers and buyers who keep seeing “Fable 5.5” in their feed and need to decide whether to wait, re-plan a migration, or ignore it. I go through eighteen days of leaks in order, grade each one by what it actually proves, and list the four public signals that would turn a rumor into a model you can call. I have no access to any unreleased Anthropic model. Everything below comes from public posts, public pages, and requests I made to public APIs on October 20.

Key takeaway

• As of 03:13 UTC on October 20, 2026, Anthropic has not announced Claude Fable 5.5. Its models overview still lists four current models: Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5.
• The strongest leak, an Azure catalog row named claude-fable-5-5 that showed up on October 13, was later deleted from LiteLLM. Its price and context numbers look like copies of the Fable 5.1 profile.
• Two client-side artifacts are harder to wave away: a Fable 5.5 special case in Cursor 3.25.6 and the claude-onigiri-eap codenames. Neither one gives a date, a price, or a confirmed mapping to Fable 5.5.
• Until there is an official row with a model ID and a rate card, plan around Opus 5.5 ($4/$20 per million tokens) and Fable 5.1 ($10/$50). Treat every Fable 5.5 spec in circulation as a placeholder.

Where the Claude lineup stands on October 20

You need the baseline to read any of the leaks. Anthropic shipped Claude Fable 5.1 on September 1. On September 22 it introduced Claude Opus 5.5 as “the first model in our new Claude 5.5 family,” and said Opus 5.5 performs at the level of Fable 5.1 on most work at 40% less than Opus 5. Sonnet 5.5 followed in late September and Haiku 5.5 on October 7. So Fable is the only line on the current models page without a 5.5 version. That gap explains why “Fable 5.5” feels inevitable, but a gap is not evidence.

Anthropic’s models overview, captured October 20, 2026. Four current models are listed, and the newest Fable API ID is claude-fable-5-1 ($10/$50 per million input/output tokens). The page tells most workloads to start with Opus 5.5. There is no Fable 5.5 row.

The same page also sets the floor for any price speculation. Fable 5.1 lists at $10 input and $50 output per million tokens. Opus 5.5 is $4/$20, Sonnet 5.5 is $2/$10, and Haiku 5.5 starts at $0.10/$0.50. All four list a 1M-token context window, 128K max output, and a June 2026 reliable knowledge cutoff. That last detail matters later, because several leaks rely on what a model “knows.”

Eighteen days of leaks, in order

The thread is long. Here are the moves that shaped the conversation, with the source each claim depends on.

Date (2026)	What surfaced	Source	What it can prove
Oct 3	Claim that Fable 5.5 appears for some accounts when Fable 5.1 is selected	@Avenoxai	Nothing about model identity; a routing feel
Oct 6	A viral motion-design prompt “for Opus 5.5 & Fable 5.5” (about 5,900 likes)	@0xMovez	The name was already in circulation as if it shipped
Oct 13	claude-fable-5-5 row in an Azure catalog, found through a LiteLLM pricing sync, with “GenerallyAvailable” status and a December 5, 2027 retirement date	@themacrosift, as compiled by Kie.ai	A string existed in a partner data feed
Oct 13	Post saying Fable 5.5 is “GA on Azure” with 1M context, 128K output, $10/$50	@xiax0603	The specs repeat Fable 5.1’s profile
Oct 15	claude-onigiri-eap in Artificial Analysis’s frontend, with low to xhigh effort and an Opus 4.8 fallback	@LuminaBench, reposted by @aGim_asf	An early-access codename exists
Oct 16	LiteLLM removes the Azure entry; a test request returns 404 DeploymentNotFound	@Mr_Salio, as compiled by Kie.ai	The Azure row did not point at a live deployment
Oct 16	”VoxelBench” knowledge probe and a claim that Opus 5.5 traffic is routed to Fable 5.5	@chetaslua	Nothing about identity (tested in our companion post)
Oct 17	Cursor 3.25.6 handles fable-5.5 / fable-5-5 IDs separately from Fable 5	@aitrackerbot, as compiled by Kie.ai	A client prepared for the ID
Oct 17	”Verrine” and “Onigiri” in the Devin CLI binary; claude-onigiri-eap-2 on Artificial Analysis; claims that partners have access	@AiBattle_, @mark_k	More codenames; partner access remains a claim
Oct 19	Blender Tesla Roadster “from fable 5.5”; “second chkpt. delivered to partners yesterday”	@chetaslua	One person’s report of a private channel

Two things stand out once the dates are lined up. First, the name “Fable 5.5” was being used as if the model had shipped by October 6, a week before the first infrastructure artifact. A lot of the later excitement is people finding evidence for a name they already believed in. Second, nearly every hard artifact comes from a partner or third-party surface (Azure, LiteLLM, Artificial Analysis, Cursor, Devin). None comes from Anthropic. That is what you would expect if a launch is being staged. It is also what you would see when a lot of people are scraping a lot of config files.

Grading the evidence, strongest first

I sort these leaks into three groups, and the boundary I use is whether a stranger could check the claim.

Checkable, but already weakened: the Azure row. A model string inside a cloud catalog is the kind of artifact that has come before real launches. Kie.ai notes that Opus 5.5 went from its first Azure catalog sighting to general availability in about two days. This row, though, carried the same 1M context, 128K output and $10/$50 price as Fable 5.1. The LiteLLM entry was then removed after an Azure Models API check found no matching model, and a test request returned DeploymentNotFound. That doesn’t prove the row was an error. It does mean the row can’t carry the claim that weights were already sitting on Azure, and it makes the circulating specs look like inherited placeholders. Anyone quoting “$10/$50, 1M context” for Fable 5.5 is quoting Fable 5.1.

Checkable and still standing: client code and codenames. A special case for fable-5-5 in Cursor’s stable build is product code a company chose to ship, not an automated sync. Two independent sightings of the Onigiri codename (Artificial Analysis’s frontend and the Devin CLI) are also stronger than one. But a codename isn’t a product name. The popular mapping from @fanofaliens says Beignet was Sonnet 5.5, Verrine was Haiku 5.5 and Wafer was Opus 5.5, so Onigiri must be Fable 5.5 or Mythos 5.5. I haven’t seen Anthropic confirm any of those pairs. Even the person who posted it ends with a question mark.

@mark_k’s October 17 post, captured October 20 (189 likes, 14.6K views at capture). The post itself says “Whether Onigiri is actually Fable 5.5 remains unconfirmed, and the suggested launch dates are guesses.” The screenshot shows what was reported, not that the codename maps to Fable 5.5.

Uncheckable: routing feels, demos and partner deliveries. “Some Opus 5.5 chats are being routed to Fable 5.5” comes back every few days. From outside, it can’t be falsified. The serving stack around a model can change without the weights changing, and a model’s answer about itself is just generated text. I tested the VoxelBench probe that drove the October 16 version of this claim, with controls, on the shipped Fable 5.1 and Opus 5.5. The shipped Fable 5.1 said “YES, I know VoxelBench” in one of three tries. The details are in the companion post. The probe produces false positives on a model that has been public since September 1.

@chetaslua’s October 19 post, captured October 20 (172 likes, 8,321 views at capture). It attributes a Blender render to “fable 5.5” and says a second checkpoint reached partners. It names no partner, model ID, or client, so the screenshot proves only that the claim was made.

The October 19 post is still the most interesting one in the thread. A checkpoint delivered to partners is a real step that happens before launches, and it leaves traces on partner surfaces. But look at what the claim says when you read it literally. If a build went to partners on Sunday, a public launch had not happened by Sunday. “Better than first” has no task, metric, or comparison attached. The Roadster, meanwhile, is Blender driven from Python, which is a long-documented interface, and a car is mostly boxes, wheels and smooth curves. A good render tells you something about capability. It tells you almost nothing about which model made it.

What we checked on October 20

I wanted my own dated record, not a repost of someone else’s, so I checked four surfaces this morning (UTC).

1. Anthropic’s models overview lists four current models, as shown above. No Fable 5.5 row.
2. The rumored docs path platform.claude.com/docs/en/models/fable-5-5/overview returns HTTP 200, which some posts read as “the page is live.” The rendered page is a generic “Page not found” that links to Fable 5.1 and Opus 5.5. A 200 status on a not-found page is a soft 404, so the status code proves nothing either way.
3. Anthropic’s site returns 404 for anthropic.com/claude-fable-5-5.
4. SandBase’s API lists anthropic/claude-fable-5.1 and anthropic/claude-fable-5 in /v1/models. A chat request to anthropic/claude-fable-5.5 returns HTTP 404 with model_not_available and the message “model is coming soon and not yet available for API calls.”

The rumored models/fable-5-5/overview URL, captured October 20. The server answers HTTP 200, but the page says “Page not found” and points to Fable 5.1 and Opus 5.5. Earlier posts reported a 404 here. Either way, the path shows no Fable 5.5 content.

The SandBase result needs careful reading. A reserved model page means SandBase has listed the ID ahead of time so a launch-day route can open quickly. It is not evidence about Anthropic’s timing, and SandBase has no early access to the model.

SandBase’s anthropic/claude-fable-5.5 page, captured October 20. Status and pricing read “Coming soon,” context window reads N/A, and the page says requests return 404 model_not_available until the provider makes the model generally available. The 128K max-output figure shown is a placeholder, not a confirmed spec.

Why the timing keeps slipping

The thread has called at least four dates: October 15, October 17, the October 21 Polymarket threshold, and today. Kie.ai reported that Polymarket priced a new Fable by October 17 at 47.5% and by October 21 at 79.5%. The October 17 contract was set to resolve no. A prediction market measures what traders expect. It doesn’t measure what Anthropic is doing. And because traders read the same leak accounts, the market partly echoes the rumor back.

Two forces really do push the timing around. The first is the 5.5 family pattern. Opus, Sonnet and Haiku each moved to 5.5 within about 15 days, so a Fable 5.5 is the obvious next step. The second is Anthropic’s own posture. The Opus 5.5 announcement calls it the company’s “first release since we called for pacing the frontier” and says external evaluators tested it before release. A Fable-tier model would presumably go through at least that much review. Public posts can’t show how long that takes.

What would actually confirm a launch

Every Claude launch this year has moved several public surfaces together. When Fable 5.5 is real, look for these four:

• A Fable 5.5 row in the models overview with an API ID, plus a model page that renders real content.
• A rate card. The price is the fact most likely to differ from Fable 5.1, so treat any number until then as a guess.
• An Anthropic announcement or release-notes entry, ideally with a system card.
• Live routes on partner clouds (Bedrock, Vertex AI, Foundry) and on aggregators like SandBase, where a request returns a completion and not model_not_available.

A catalog row alone, a docs URL alone, or a demo alone has fired early before. All four at once is a launch.

What to do while it is still a rumor

Don’t hold production work for a model that has no ID or price. Anthropic’s own guidance is to start most workloads on Opus 5.5 and move to Fable 5.1 for “demanding reasoning and long-horizon agentic work” or when Opus 5.5 at higher effort falls short. At list prices, a request with 1M input tokens and 200K output tokens costs $20 on Fable 5.1, $8 on Opus 5.5, and $4 on Sonnet 5.5. A Fable-tier upgrade has to beat a 2.5x price gap on your tasks, not just look good in demos.

Three cheap preparations make launch day boring:

• Keep model IDs in config. Hard-coded claude-fable-5-1 strings are the main reason a launch-day evaluation turns into a code change.
• Freeze an evaluation set now. Pick 20 to 50 real tasks, record success, tokens, latency and rework on Fable 5.1 and Opus 5.5 today, and you’ll have a baseline the day a new ID appears. Our Fable 5.1 long-running agent post describes one way to set this up.
• Watch, don’t poll the feed. The models overview and partner catalogs will move before or alongside the announcement. Viral demos don’t change your decision.

FAQ

Is Claude Fable 5.5 released?

No. As of October 20, 2026 (03:13 UTC), Anthropic has not announced it. Its models overview lists Fable 5.1 as the newest Fable model, and the rumored docs page renders “Page not found.”

What is Claude Fable 5.5 pricing and context window?

There’s no official number. The circulating 1M context, 128K output and $10/$50 per million tokens match Fable 5.1’s published profile. They appear to have been copied into the Azure catalog row that was later removed.

Is claude-onigiri-eap Fable 5.5?

Unconfirmed. The codename appears in Artificial Analysis’s frontend and, per community reports, in the Devin CLI. Anthropic has not mapped it to a product. The Beignet/Verrine/Wafer pairings people cite are community inference.

Can I call Fable 5.5 on SandBase?

Not yet. anthropic/claude-fable-5.5 returns 404 model_not_available on October 20. You can call anthropic/claude-fable-5.1 and anthropic/claude-opus-5.5 today. See the Anthropic vendor page, or read how Fable 5.1 differs from Mythos 5.1 in access.

Is my Opus 5.5 chat secretly running Fable 5.5?

Nothing you can run from outside will tell you. Self-reported identity and knowledge probes are generated text, and our controlled probe got a false “yes” from the shipped Fable 5.1. Use the model string returned by the API as your contract.

Bottom line

The Fable 5.5 trail is real in one narrow sense: several partner and client surfaces have prepared for a fable-5-5 ID, and at least two codenames are in testing. It is not a release. The specs in circulation are inherited from Fable 5.1, the strongest artifact was withdrawn, and the newest claim is about a private partner channel. The useful posture this week is the boring one: ship on Opus 5.5 or Fable 5.1, keep your model ID configurable, and wait for a row with a price. I’ll update this page when Anthropic publishes one.

Sources checked October 20, 2026 (UTC): Anthropic models overview and Opus 5.5 announcement; the rumored Fable 5.5 docs URL; SandBase /v1/models, /v1/chat/completions and model page; X posts linked above; compilations by Kie.ai, OrcaRouter and Crypto Briefing. Items marked “as compiled by” were not independently verified.