Claude Fable 5.5 Leak: What Is Verified So Far

The short version: a Fable 5.5 rumour has been building since August, peaked on 1 October, and remains unverified. The strongest evidence is circumstantial: screenshots in which the interface still labels the model Fable 5.1 while the poster says the output felt different. The weakest is anything where a model is asked what it is.

• Official status: no announcement, no system card, no model page. Anthropic's documentation still identifies Fable 5.1 as the current Fable.
• Confirmed releases: Opus 5.5 launched on 22 September 2026 and Sonnet 5.5 followed days later. Those are separate products from Fable.
• Leak claims: a fresh pretraining run, possible silent routing of some users, and a launch window of late September to early October.
• Demos: several clips (an animated X-ray scene, a cartoon, a voxel floating-island world and a 3 minute 40 second animation called "The Grader") are circulating, with unverified attribution.
• Benchmarks: none published or credibly claimed.

Our view: a new Fable model in the next few weeks is plausible given the cadence, but plausibility is not evidence. Plan for the possibility, do not rebuild anything around it.

The release history gives context. Fable 5 shipped on 9 June 2026, Fable 5.1 on 1 September, and Opus 5.5 on 22 September, with Sonnet 5.5 and a Haiku 5.5 announced alongside or shortly after. Our Opus 5.5 leak scorecard shows how an earlier Anthropic leak performed: it matched the price and launch date but its Fable 5.2 element never shipped under that name.

Notice the renaming. The same pretraining story carried two version numbers within five days. That fits a rumour that tracks a real underlying event, but it also fits one that adapts its label as the naming scheme changes. After Opus 5.5 pulled the family to 5.5, calling the next Fable "5.5" was the natural guess.

It helps to list the claims that can be checked against primary sources, and those that cannot.

Verified

• Anthropic's documentation and model directory list Fable 5.1 as the current Fable. Fable 5.5 is not listed.
• Opus 5.5 is real and released, covered in our Opus 5.5 review. Sonnet 5.5 is also real, covered in our Sonnet 5.5 review.
• Fable 5.1's published specifications, as summarised by reviewers: 1 million token context, 128,000 token output, price of $10 per million input tokens and $50 per million output tokens (around £7.40 and £37), knowledge cutoff of June 2026.

Not verified

• That any model called Fable 5.5 exists.
• That users are being silently routed to a newer model.
• That any demo clip was generated by a model newer than Fable 5.1.
• Any performance claim, since no benchmark has been published.

A practical rule: if the claim cannot be matched to an Anthropic page, an API model identifier or a system card, it is a rumour.

The first-preview roundups centre on a handful of creative clips. WorldofAI's video shows examples covering animation, 3D environments, coding and storytelling, as its description says. The clips reported in the wider coverage include an animated X-ray scene of roughly 15 minutes, a cartoon with a stated prompt, a voxel floating-island world and "The Grader", a 3 minute 40 second animation reported to use Blender and the ElevenLabs API.

Why demos prove little

1. Selection bias. People share the best of many attempts.
2. Tooling. A clip that pipes through Blender and a voice API tells you about the pipeline and the person directing it as much as about the model.
3. No control. Without the same prompt run on Fable 5.1, nobody can say whether the clip is better than what the current model produces.
4. Attribution. The reporting we reviewed describes the attribution of these clips as unverified.

WorldofAI says it benchmarked what it shows using its own tool, WoAI Bench. We have not seen published scores for Fable 5.5 from that tool, and the video is a roundup that also covers Gemini 4 Argon and Claude Code, so check the video itself for details. For the Google release mentioned there, see our Gemini 4 Argon review.

The most interesting claim from 1 October is that some users are being silently routed to a newer model while the interface keeps reporting Fable 5.1. The evidence is two screenshots, one from the Claude website and one from Claude Code, both still labelled Fable 5.1.

This is plausible in principle. Labs commonly test models on a small share of traffic before launch, and WorldofAI's video discusses possible early auto-routing. But a screenshot with a Fable 5.1 label proves only that the label was Fable 5.1. It says nothing about the model behind it. A reviewer's careful summary is that identification requires independently verifiable information, not recall tests or output quality alone.

The self-report trap

Asking a model what it is, or what its knowledge cutoff is, produces answers that depend on its training data and system prompt. A model may not know its own version number. It may repeat the name in its system prompt. It may guess. None of those outcomes tells you which model is running.

If you build with Claude, you do not have to rely on screenshots. This checklist works for any "secret model" rumour.

1. Check the API model identifier. The `model` field in an API response is what you were actually served. An official release will have a documented identifier; a rumour will not.
2. Check the docs. Anthropic's model overview, pricing page and the supported-model list in Claude Code are the sources of truth.
3. Run a fixed prompt set. Pick 10 to 20 tasks that matter to you and run them on the current model, saving outputs. Re-run when a new model is announced, and compare.
4. Score blind. Have someone remove model names before you compare outputs.
5. Log dates and versions. If behaviour changes without an announcement, a dated log tells you when.
6. Distrust round numbers. Rumoured benchmark figures with no source are usually invented.

Our best AI models guide and frontier benchmark roundup show how to compare models when official numbers do arrive.

Fable 5 shipped on 9 June and Fable 5.1 on 1 September, about twelve weeks apart. A Fable 5.5 on 1 October would have arrived after roughly 30 days, the fastest turnover yet. That is not impossible, because the Opus 5.5 launch on 22 September shows the company moving quickly. It does cut against the pattern.

Two competing explanations fit the facts:

• A real pretraining run is nearly ready. The August and September posts describe it, and labs often test before launch.
• Naming is being projected onto a normal point release. After Opus 5.5, any new Fable would be expected to carry a 5.5 label whatever its actual version.

Neither can be ruled out. What we can say is that earlier Anthropic rumours, including those assessed in our Opus 5.5 leak scorecard, were right about some things and wrong about others, usually the naming.

The timing of the rumour overlaps with Anthropic's reported IPO preparations. One social-media post argued that Anthropic will "build maximum momentum" ahead of a November listing and that a new Fable is how it does so. That is opinion, not evidence. But the business context is real: the draft prospectus, covered in our Anthropic IPO analysis, describes huge compute commitments that the company needs to keep filled with paying usage. Strong model releases fill capacity. It would be surprising if the commercial calendar and the product calendar were unrelated, though that does not tell us when any specific model will ship.

For the legal and policy context around the company, see our explainer on the US government blacklist.

This is not the first time. Our archive covers the Fable 5.1 release-date speculation, the period when Fable 5 was pulled, its return and the question of when it would come back, plus the comparison in Fable 5 vs Mythos 5. Three lessons recur.

1. Dates are the best-sourced part of a leak, names the worst. Expect a model to appear on roughly the predicted timeline under a slightly different name.
2. Capability claims inflate. "Beats everything" claims before release rarely survive contact with independent testing.
3. Silence from the lab is normal. Anthropic does not comment on rumours, so absence of denial means nothing either.

• Do not change production models on a rumour. Pin a model identifier and change it deliberately.
• Prepare your evaluation set so you can compare quickly when a real release lands.
• Budget for the premium tier. If Fable 5.5 follows Fable 5.1's price of $10 and $50 per million tokens, that is about £7.40 and £37, but pricing for an unannounced model is unknown.
• Consider cheaper options. Sonnet 5.5 is already available and, per Julian Goldie's coverage this week, is pitched as faster and cheaper than its predecessor.
• Follow primary sources. Anthropic's news page and model documentation beat any aggregator.

We cannot compare an unreleased model. We can note the competitive picture it would land in. OpenAI's GPT-6 Astra and GPT-6.1 Sol are current (see our Astra review and GPT-6.1 Sol coverage), Google launched Gemini 4 Argon on 30 September, and our Opus 5.5 vs Astra comparison sets out where Anthropic's current models stand. If a Fable 5.5 arrives, expect the benchmark comparison to be the first thing independent testers publish.

Understanding how labs release models makes the silent-routing claim easier to judge. A new model rarely goes from lab to everyone in one step. The usual stages are internal testing, a small share of live traffic for a subset of users, a wider trial, and finally a public announcement with documentation and an API identifier.

What a staged rollout looks like from outside

During the traffic-trial stage, a few users might get a different model with an unchanged label. If that happened, you would expect scattered reports of output that felt different, an inability to reproduce it consistently, and no official statement. That is exactly what the 1 October posts describe. The pattern fits, but a pattern is not a confirmation. A model changing between sessions can also come from ordinary variation, different system prompts, tool use, or a user's own prompt changing.

Auto-routing is a separate idea

WorldofAI's description also mentions possible early auto-routing. Routing means a system picks the model per request, for example sending hard tasks to a larger model. Microsoft's new Copilot offers an auto-router across vendors, as we describe in our Copilot Autopilot guide. A routing layer inside a model provider would be a product decision, and one that Anthropic would normally document, because developers need predictable behaviour.

Why labs do this

Real traffic exposes failure modes that internal evaluations miss, such as odd tool-calling patterns or long-context edge cases. It also gives a cost and latency picture. None of this is evidence for or against the Fable 5.5 claim. It only explains why the claim is believable.

The best defence against hype is a small private test. You do not need a research lab to do it.

What to include

1. Five tasks you do every week, such as drafting a client email, summarising a report or fixing a bug.
2. Three hard tasks where current models usually fail, such as multi-step reasoning with a trap in it.
3. Two long-context tasks, such as answering questions about a long document you know well.
4. Two agentic tasks, such as a multi-file code change or a research task that needs several sources.

How to run it

Save the exact prompts. Run each on the current model and store outputs with the date and model identifier. When a new model is announced, run the identical set and compare. Score on three things: correctness, how much editing the output needs, and cost per task. A model that is slightly better but twice the price may not be better for you.

How to read results

One run is anecdote. Repeat each task a few times, because language models vary. If the new model wins consistently on your tasks, it is worth adopting. If the difference is within the noise, wait for the price to fall. Our AI API pricing comparison and best AI coding agents guides can help you set a baseline.

Creator channels are a fast, useful signal. They are also commercial. The WorldofAI description includes a sponsor link and promotes a community and a benchmarking tool. That does not make the reporting wrong, but it is a reason to separate three things: what the creator saw, what the creator inferred, and what the creator was told by others.

• Saw: clips on screen, screenshots, a model's interface.
• Inferred: that a clip came from a newer model, or that routing is happening.
• Reported: claims from X accounts the creator cannot verify.

Good practice is to follow a claim back to its first source. In this case the earliest posts are from individual accounts without documentation, and the later posts cite the earlier ones. A rumour repeated by ten channels is still one rumour. For a model we can actually review, see Claude Opus 5.5 and the Anthropic IPO prospectus coverage, both grounded in official or widely reported primary material.

There is a real chance that Anthropic is testing a newer Fable, and the evidence for it is weak. Screenshots with an unchanged label, unattributed demos and unsourced timelines add up to a rumour, not a release. The sensible response is neither dismissal nor excitement. Keep an evaluation set ready, watch Anthropic's own pages, and treat any benchmark figure that arrives without a source as noise. We will update this article if and when the company confirms anything.