What Is Claude Fable 5.5?
At the moment, “Claude Fable 5.5” is best understood as an unconfirmed model name rather than a released Anthropic product.
Anthropic’s current model documentation lists Fable 5.1 alongside the released Haiku, Sonnet, and Opus models. It does not contain a Fable 5.5 entry, API model string, pricing table, system card, or release announcement.
That distinction matters. A model can exist internally without becoming a public product. Anthropic may test several checkpoints, route limited traffic through them, rename them before release, or decide not to release them at all.
Even the version name is uncertain. The assumption that Anthropic must move from Fable 5.1 to Fable 5.5 is based on the version numbers used by other Claude tiers. It is not an announced roadmap.
In community discussions, “Fable 5.5” has become a convenient label for several separate observations:
A model name briefly appeared in Azure-linked catalog data.
Users reported changes in Claude’s coding and Artifact outputs.
An unidentified Claude checkpoint appeared in public benchmark-site code.
Creators attributed complex SVG, Three.js, and browser-game outputs to a new Fable model.
Those observations may point to ongoing Anthropic testing. They do not prove that every reported output came from the same model—or that the final public name will be Fable 5.5.
In plain English: there are clues, but no product announcement.
Is Fable 5.5 Real? Confirmed Facts vs Rumors
The evidence becomes easier to understand when each claim is assigned a confidence level.
| Claim |
Evidence level |
What it means |
A claude-fable-5-5 name appeared in Azure-linked catalog data |
Confirmed event |
The name briefly existed in synchronized third-party data |
| Anthropic has announced Fable 5.5 |
False as of October 19 |
No official announcement or model entry exists |
| Fable 5.5 costs $10 input and $50 output per million tokens |
Unsupported |
Those values were copied from Fable 5.1 |
| It has a 1M-token context window and 128K output limit |
Unsupported |
Those fields were also inherited from Fable 5.1 |
| Anthropic is testing an unreleased Claude checkpoint |
Plausible |
Public code contains signs of at least one unidentified checkpoint |
| That checkpoint is definitely Fable 5.5 |
Speculation |
No public evidence connects the two names |
| Some Claude users have already received Fable 5.5 |
Unverified |
Users cannot inspect the actual checkpoint selected by the routing layer |
| Fable 5.5 is stronger than Opus 5.5 or GPT-6 Astra |
Unknown |
No reproducible head-to-head evaluation exists |
The most defensible conclusion is narrow: Anthropic appears to be testing unreleased Claude systems, and the Fable 5.5 name briefly surfaced in catalog data. Everything beyond that needs qualification.
The Fable 5.5 Leak Timeline
The timeline matters because many summaries captured the original leak but missed what happened next.
October 13: The Azure/LiteLLM Entry Appeared
On October 13, an automated synchronization process added azure_ai/claude-fable-5-5 in LiteLLM PR #46201.
That sounds convincing until the fields are inspected.
The model name and retirement date reportedly came from Azure Models API data. The remaining fields—including the $10/$50 token rates, 1M-token context window, 128K maximum output, and capability flags—were inherited from Fable 5.1.
Inherited fields are useful placeholders for maintaining a model catalog. They are not an Anthropic specification sheet.
Therefore, the entry established one fact: an Azure-linked data source briefly returned the Fable 5.5 name. It did not establish the model’s price, context window, output limit, release date, or public availability.
October 14: The Entry Was Removed
The following day, LiteLLM PR #46428 removed the model.
The follow-up audit checked several Azure regions, including East US, East US 2, West US 3, and Sweden Central. None exposed a usable Fable 5.5 deployment. An attempted request returned 404 DeploymentNotFound.
This changes the interpretation of the original leak. The catalog entry was real, but it did not lead to a callable public model.
It may have been premature provider metadata, a staged listing, a temporary catalog error, or a deployment that never became externally available. The public evidence cannot distinguish among those explanations.
October 15: “Onigiri” Appeared
Another clue appeared when an unreleased identifier, claude-onigiri-eap-2, was found in public-facing benchmark-site code. The configuration included Low, Medium, High, and Xhigh effort levels and used an existing Claude model as a fallback.
This is a stronger technical signal than a user saying, “Claude felt smarter today.” It indicates that somebody prepared a configuration for an unreleased Claude checkpoint.
Still, the code did not contain the name Fable 5.5. Onigiri could represent a Fable test, an Opus test, a routing experiment, or an internal checkpoint that never receives a public release.
Onigiri = Fable 5.5 remains a hypothesis.
Fable 5.5 Release Date: What Do We Actually Know?
Anthropic has not announced a Fable 5.5 release date.
Predictions from social accounts and community posts should not be converted into a calendar date. The Azure-linked entry did not include an official launch announcement, and its removal makes an immediate public deployment less certain—not more.
| Signal |
Does it confirm a release date? |
| Anthropic announcement |
Yes, but none exists |
| Official model documentation |
Yes, but no Fable 5.5 entry exists |
| Published system card |
Yes, but none exists |
| Azure catalog appearance |
No |
| LiteLLM model entry |
No |
| Community routing report |
No |
| Social-media prediction |
No |
An internal checkpoint could reach production under the Fable 5.5 name. It could also be renamed, held for more evaluation, merged into another Claude release, or abandoned.
My judgment is that the technical traces make ongoing testing plausible. They do not support a specific release date.
Fable 5.5 Pricing and Features Are Not Confirmed
There is no verified Fable 5.5 pricing.
The frequently repeated $10 per million input tokens and $50 per million output tokens belongs to the current Fable 5.1 pricing reference used in the synchronized entry. It was not a new pricing announcement.
The same problem applies to the rumored 1M-token context window and 128K output limit. Those numbers may eventually turn out to be correct, but the leaked catalog record cannot prove it because the fields were copied from an earlier model.
Fable 5.1 can serve as a rough indication of what the existing Fable tier costs. It should not be presented as a Fable 5.5 quote.
What Features Have Users Reported?
Community-attributed examples tend to cluster around several tasks:
SVG illustrations and animations;
Three.js scenes;
browser games;
long interactive Artifacts;
complex interfaces generated from short instructions.
A community-curated collection of alleged Fable 5.5 outputs includes several such projects. They are useful as clues about the kind of frontend and visual-coding work Anthropic may be testing.
They are not controlled benchmarks.
The creators generally cannot retrieve a private serving checkpoint ID. Their attribution may be based on timing, perceived output quality, or an unverified routing theory. Prompt selection and failed attempts may also be missing from the published examples.
The safe conclusion is that users have reported unusually complex outputs and attributed them to Fable 5.5. It is not yet safe to claim that Fable 5.5 has a particular capability, success rate, or advantage over another model.
Is Fable 5.5 Available in Claude Code or via API?
No public Fable 5.5 model ID is currently listed for the Claude API or Claude Code.
The official Claude Code model configuration does not include claude-fable-5-5. The Azure-linked identifier also failed availability checks before being removed.
That means developers should not:
place an unconfirmed Fable 5.5 string in production configuration;
advertise Fable 5.5 support based on a catalog screenshot;
assume a successful Claude request was silently routed to Fable 5.5;
ask Claude to identify itself and treat the answer as server metadata.
A model’s prose cannot authenticate the backend that served it. The reliable evidence is the model identifier returned by the API or documented by the provider.
How to Verify a Currently Available Claude Model
Until Fable 5.5 receives a public model string, test with a released model such as Sonnet 5.5. GPT Proto documents the OpenAI-compatible base URL as https://gptproto.com/v1 and the Sonnet model string as claude-sonnet-5-5. gptproto.com
Store the API key in an environment variable:
export GPTPROTO_API_KEY="your_api_key"
Then send a basic request:
curl --fail-with-body \
--request POST "https://gptproto.com/v1/chat/completions" \
--header "Authorization: Bearer $GPTPROTO_API_KEY" \
--header "Content-Type: application/json" \
--data '{
"model": "claude-sonnet-5-5",
"messages": [
{
"role": "user",
"content": "Reply with the word ready and nothing else."
}
]
}'
--fail-with-body makes curl return an error for unsuccessful HTTP responses while preserving the response body for diagnosis.
Check the response metadata rather than asking the assistant which model it is. Also confirm the live identifier in the relevant model page before deployment because availability can change.
You can use the same request structure with the published IDs shown on the Claude Haiku 5.5 API, Claude Sonnet 5.5 API, and Claude Opus 5.5 API pages.
Do not substitute claude-fable-5-5 until a live model page or official documentation confirms it. Guessing model strings is not model discovery.
Why the “Tibo Test” Cannot Identify Fable 5.5
One community theory proposes turning off web search and asking Claude about a specific person, phrase, or inside joke. If Claude recognizes it, the theory says, the user has been routed to Fable 5.5.
The community discussion around alleged secret testing is interesting as a record of how users investigate opaque model updates. It is not a reliable identification method.
There are four problems.
First, an answer is not a model ID. A language model can state the wrong identity, repeat a name supplied in the prompt, or infer what the user expects.
Second, the underlying model is not the only variable. System prompts, retrieval, knowledge patches, routing rules, and product-level tools can all change a response.
Third, recognition tests are vulnerable to contamination. Once a phrase spreads through public posts, newer retrieval layers or future training data may contain it.
Fourth, the test lacks controlled reproduction. Users cannot guarantee identical checkpoints, system instructions, parameters, accounts, or routing conditions.
The Tibo test is a rumor-tracking ritual. It is not an API-level model detector.
Fable 5.5 vs Haiku 5.5, Sonnet 5.5, and Opus 5.5
There is no defensible Fable 5.5 benchmark comparison yet. Any table assigning it a context window, score, speed, or quality ranking would mix an unverified model with released products.
A useful comparison should instead answer a practical question: wait for Fable 5.5, or deploy something available today?
| Model |
Publicly available |
Current role |
Practical decision |
| Fable 5.5 |
No |
Unconfirmed |
Do not build production plans around it |
| Haiku 5.5 |
Yes |
Classification, extraction, routing and high-volume sub-agent steps |
Choose when throughput and cost matter |
| Sonnet 5.5 |
Yes |
Coding, document work and scoped agent workflows |
Best balanced option for most applications |
| Opus 5.5 |
Yes |
Complex coding and open-ended, long-running work |
Choose when task difficulty matters more than cost |
Fable 5.5 vs Haiku 5.5
Haiku 5.5 is designed for a different job from the rumored Fable model. It handles classification, extraction, summaries, routing, customer-support steps, and narrow sub-agent tasks where request volume matters.
Its official pricing uses two prompt-length tiers. Prompts up to 100K tokens cost $0.10 per million input tokens and $0.50 per million output tokens. When a request exceeds 100K input tokens, the full request moves to the $0.50 input and $2.50 output tier.
Fable 5.5 might eventually offer more reasoning capacity, but that would not automatically make it a sensible replacement for a cheap routing or extraction model.
For high-volume workloads, use the released Claude Haiku 5.5 API instead of waiting for an unpriced flagship.
Fable 5.5 vs Sonnet 5.5
Sonnet 5.5 is the most practical default for developers who need coding, bug fixing, document analysis, visual understanding, or bounded agent workflows now.
Anthropic’s direct list rate is $2 per million input tokens and $10 per million output tokens. More importantly, it has a documented model ID and can be evaluated in your own application.
A future Fable model may outperform it on the hardest tasks. The likely trade-off would be cost and possibly latency. Until both models can be tested using the same prompts and settings, that trade-off cannot be quantified.
If your project is ready to ship, use Claude Sonnet 5.5. Waiting for an unannounced model is not an evaluation strategy.
Fable 5.5 vs Opus 5.5
Opus 5.5 is the closer practical comparison because it targets complex coding, longer tasks, and work requiring sustained judgment.
Its official direct rate is $4 per million input tokens and $20 per million output tokens. Unlike Fable 5.5, developers can call it, measure it, and decide whether the extra task quality justifies the bill.
There is no public benchmark showing that Fable 5.5 beats Opus 5.5. Community demos may suggest where Anthropic is experimenting, but they do not provide matched prompts, repeated trials, latency measurements, or failure rates.
For difficult production work today, evaluate the Claude Opus 5.5 API.
Is Fable 5.5 Better Than GPT-6 Astra?
Nobody can answer this with reliable evidence yet.
A valid comparison requires an identified Fable 5.5 endpoint, fixed model versions, identical prompts, disclosed settings, repeated trials, and scoring criteria suited to the task. None of those conditions currently exists.
Any confident winner is speculation dressed as a benchmark.
Should Developers Wait for Fable 5.5?
Most developers should not wait for Fable 5.5.
Choose according to the work you need to run:
Use Haiku 5.5 for classification, extraction, summaries, routing and repeated sub-agent steps.
Use Sonnet 5.5 for everyday coding, bug fixes, documents and bounded agent workflows.
Use Opus 5.5 for complex coding, long-running tasks and open-ended knowledge work.
Evaluate Fable 5.1 only when your internal tests show that cheaper models cannot complete the task reliably.
Monitor Fable 5.5 if you research frontier models, but do not make delivery dates depend on it.
A future Fable release could become the better choice for a small set of difficult workloads. That possibility does not make it deployable today.
Compare the live rates and currently supported models on the GPT Proto pricing page rather than budgeting around leaked placeholder data.
What Would Confirm a Real Fable 5.5 Release?
The status should change from “rumored” to “confirmed” only when at least one authoritative artifact appears:
An Anthropic release announcement;
an entry in Anthropic’s official model documentation;
a public API model ID;
an official pricing table or system card;
a Claude Code configuration entry;
or a reproducible API response from a documented endpoint.
A social post is not enough. Neither is a catalog screenshot, a model claiming to be Fable 5.5, or one unusually good Artifact.
Once a public endpoint exists, the next questions should be empirical: What does it cost per completed task? How often does it succeed? How long does it take? Which workloads improve enough to justify switching?
Until then, the accurate answer is less exciting but more useful: Fable 5.5 may exist as a test checkpoint, but its public product details remain unconfirmed.
