Claude Fable 5.5 Now Has Two Checkpoints and Still No Vendor Page

Today is Tuesday 20 October 2026, which matters only because it is the date the Claude Fable 5.5 rumour set for itself. On Thursday 15 October, Andrew Curran posted on X that he had expected the release “next Tuesday” before revising toward that same day, and the Tuesday his sentence resolved to is this one. One day before it arrived, on Monday 19 October, the X account that has been posting Fable demos all month added a new sentence to the story: the Blender render it showed last week was Claude Fable 5.5 at its first checkpoint, and a second checkpoint was “delivered to partners yesterday” and is “better than first”. Nobody has published a model ID, a rate card, a context window or a benchmark row for either checkpoint. The model you can actually call at the frontier tier remains Claude Fable 5.1, released 1 September 2026 at $10 per million input tokens and $50 per million output tokens with a 1M-token context window, and the model A​nthropic's own documentation points most workloads at is Claude Opus 5.5, from 22 September at $4 and $20. Everything about Fable 5.5 in this piece is a claim from one X account, unverified, and labelled as such throughout. The documented AI failures and unreleased-model incidents behind stories like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

Five claims in one run-on post, sorted by whether you can check them

The signal is a single paragraph written the way people write when they are excited, which makes it worth taking apart clause by clause. The useful split is not important-versus-unimportant. It is checkable-versus-not.

• “Tesla Roadster in Blender using Python Code” — a description of an artefact. Plausible on its face, and structurally the least informative thing in the post: driving Blender from Python is a documented, decade-old interface, and a car is a friendly target for procedural geometry because it is mostly boxes, wheels and smooth surfaces. You cannot confirm the render came from any particular model, and the post does not claim you can.

• “this output from fable 5.5” — an attribution. Unfalsifiable from outside. No identifier accompanies it, and the account has no stated relationship to A​nthropic that would let a reader weigh it.

• “i gave it launch video from x and it cooked” — the prior claim, restated. This is the 16 October story: footage described as a launch video, credited to the model. The link in that post did not open without an authenticated X session, and the link in this one resolves to the same kind of shortened permalink.

• “now i am thinking this was first checkpoint” — a revision of the poster's own earlier claim. Note the verb. Not “this was” but “I am thinking this was”. The post is reinterpreting its own evidence, after the fact, without new evidence attached.

• “second chkpt. delivered to partners yesterday better than first” — three assertions in nine words: a second checkpoint exists, it reached partners on Sunday 18 October, and it is an improvement. The first is invisible, the third is unmeasured, and the second is the only one that could ever leave a trace.

Today is the date the rumour named, and a checkpoint is what turned up instead

That is the part worth sitting with. A claim with a date is a claim you can grade, and this one has now been graded by the calendar rather than by an argument. Curran's post said he thought it would be Tuesday, that it might be that Thursday, that he did not know for sure, and that deployment issues could push it a week. The Tuesday in question is today. What exists on this Tuesday is a post from an unrelated account saying a second build went to partners two days ago.

Read those two sentences together and the shape of the thing becomes clear. In a launch, a model becomes available to everyone at once and the announcement is the last domino to fall. In what the post describes, a model becomes available to a small group and everybody else keeps reading about it. Those are different events, and only one of them is a release. A partner-delivered checkpoint is, by construction, the stage that exists before a release — it is what a lab hands to cloud partners so their catalogues and SDKs are ready on day one. Saying a checkpoint went to partners is close to saying, in plain language, that a launch has not happened yet.

That is not a debunk. A team that ships a checkpoint to partners is a team that is close, and this is the most concrete staging detail to appear in six weeks of Fable 5.5 claims — more concrete than a hint, a poll or an arena codename, because partner delivery is an operational step with a recipient. It is also, conveniently for anyone trying to verify it, the one step that leaves fingerprints outside A​nthropic.

A checkpoint is not a model, whatever the post implies

The word is doing a lot of unearned work in this thread, so it is worth being precise about what it means in training. A checkpoint is a saved state of a model's weights at a point during or after a training run. A single run produces many of them, at intervals, for reasons that include resuming after a crash, comparing intermediate quality, and picking the version to post-train. Two checkpoints of the same run are the same model at two moments. They can differ noticeably on some tasks and not at all on others, and neither difference is promised to you in a document.

Which is why “better than first” is the emptiest phrase in the post. Better at what, measured how, by whom, against what? There is no eval, no comparison prompt, no side-by-side, and no third party who has touched either build. The poster is comparing their own impressions of two things nobody else has seen. That may be an honest report. It is not evidence a reader can use.

What would turn a checkpoint into a model, in the sense a reader cares about? Four things, and they have arrived together on every Claude launch this year: a row in A​nthropic's models overview carrying a model ID, a rate card, a documentation page that resolves, and a line in the API release notes. Here is the state of those surfaces as of today:

• Models overview — read today, it lists four current models: Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5. Claude Fable 5.5 has no row.

• Documentation path — the Fable 5.5 overview URL returns a 404 today, the same as it did on 18 October.

• Release notes — the newest model launch in the October notes is Claude Haiku 5.5 on 7 October. Everything since is platform work: fast mode leaving research preview for Claude Opus 5.5 on 15 October, a Compliance API beta the same day, OpenTelemetry export and Managed Agents scheduled deployments on 14 October.

• Independent scoreboard — Artificial Analysis carries no Claude Fable 5.5 row today. The leaders on its Intelligence Index are Claude Opus 5.5 at 58 and Claude Sonnet 5.5 at 56, with Claude Fable 5.1 sitting at 53 in two evaluation modes.

None of those absences disproves anything. An unannounced model is unannounced by definition, and a checkpoint delivered to partners is precisely the thing that would not yet appear in any of them. But they do establish the frame: there are two checkpoints claimed and zero surfaces, and the gap between those numbers is the whole story.

The scoreboard is the point of the piece, and it is deliberately lopsided in one direction only. Checkpoint one at least produced a video somebody can look at, even if the link does not open. Checkpoint two produced a sentence. Everything a reader could act on — the price, the identifier, the benchmark — is empty on both sides, which is what it looks like when a model is being described rather than shipped.

The one surface a partner delivery can actually leak onto

If a build really reached cloud partners on Sunday, then the trace will not appear on A​nthropic's site first. It will appear where partners keep their catalogues: a model row in a cloud marketplace, a new entry in a partner's documentation, a deployment name in a console. That is exactly the shape of the 13 October sighting, when a catalogue row for Fable 5.5 turned up on a cloud marketplace ahead of any vendor announcement — a row with nothing behind it: no price, no documentation link, no way to invoke anything.

This is why the partner claim is simultaneously the most interesting and the least verifiable thing in the post. Interesting because a catalogue row is checkable and dated. Least verifiable because the poster names no partner, no console and no row. “Delivered to partners” is a claim about a private channel, made in public, with the recipient left blank.

If you want to watch this yourself, watch the partner catalogues rather than the newsroom. A model row with a price and a resolvable documentation link is a launch in progress. A row with a name and nothing else is a placeholder, and placeholders have been sitting in cloud catalogues for models that took weeks to arrive after the row appeared. Do not treat the row as the announcement; treat it as a timer.

The Roadster is the weakest evidence in the post, and that is not an accusation

A car in Blender is a strange artefact to hang a model generation on, and the strangeness is instructive. Blender's Python API is enormous, well documented, and full of examples; a language model that has read the internet has read a great deal of it. Procedural car geometry — a body from a few extruded profiles, four cylinders, a couple of materials, a camera and a light — is a task where a competent model produces something recognisable on the first or second attempt and a very good one produces something clean. The output being impressive is not in question. What it tells you about which weights produced it is close to nothing.

Add the poster's own framing and the uncertainty compounds rather than resolves. “Still thinking about this output” is the sentence of someone impressed by a result, not someone reporting a measurement. And a demo, by its nature, is the attempt that worked. Every model vendor on earth ships demos on that basis, which is why a demo is fair evidence of capability and poor evidence of identity.

The honest position is symmetric. This account has been consistently early on Fable 5.5 and consistently vague about what it is running: no identifiers, no screenshots of a model name in a client, no transcripts. That pattern is compatible with a genuine early relationship with the model, and it is also compatible with somebody who has access to something and has decided what it is. Six weeks in, the thread has produced a great deal of demonstration and not one artefact that a stranger could check.

What to do on a Tuesday when the model is still a checkpoint

The practical question this post raises is not “is the second checkpoint better” — nobody outside the partner channel can answer that, and the answer will not be published. It is “what does my stack look like on the day the answer changes”, and that question has an answer today.

Start with the arithmetic you would be doing if a Fable 5.5 arrived at Fable-line prices. On the current rate cards, a request of one million input tokens and two hundred thousand output tokens costs $20 on Claude Fable 5.1, $8 on Claude Opus 5.5, and $4 on Claude Sonnet 5.5. The Fable tier is two and a half times Opus and five times Sonnet for the same shape of call, which is the number that should decide how much of your traffic you would move on day one even if the model is everything the post says it is. A rumour about quality is not a reason to move a production path; a price ratio is a reason to move a slice of one.

That is the case automatic failover was built for. On OrcaRouter, anthropic/claude-fable-5.1 is callable now at the vendor's own list price with 0% markup, so the $10 and $50 above is what you pay rather than a reseller's version of it, and if A​nthropic changes a rate the change is live here the same day. Put an unproven model on a share of traffic behind the same OpenAI-compatible endpoint as the model you trust, and let a failure rate or a latency threshold send requests back to the proven one rather than an engineer noticing at 3am. The routing DSL lets that be a rule rather than a rewrite, so “should we move to Fable 5.5” becomes a question you can answer with your own evals in an afternoon instead of a migration you have to schedule. For cases where the honest answer is that you want two models' strengths in one call rather than one model's, model fusion is there for that.

None of this requires believing or disbelieving the 19 October post. It requires only that the cost of being ready be near zero, which is what a single endpoint and a routing rule get you.

What this Tuesday actually settled

The date the rumour named has arrived and the model has not. What arrived instead is a claim that a second build exists somewhere, held by someone the poster does not name, described as better by someone who has not measured it. That is a more specific claim than the thread has produced before, and it is still not a model you can call.

Watch the partner catalogues this week, because that is where a delivered checkpoint would surface, and watch for a price next to a name rather than a name on its own. If a row appears with a rate card and a documentation link that resolves, the checkpoint story was true and the launch is days behind it. If the thread produces a third checkpoint next week, the pattern is worth naming: a model that keeps improving in private and never arrives in public is not a model you can plan around, however good the renders look. The version of this story that costs you nothing is the one where you keep calling Claude Fable 5.1 or Claude Opus 5.5, watch the surfaces, and swap a routing rule the day there is something to swap to.