Claude Fable 5.5: The Reference Page for a Model That Has Not Shipped

Claude Fable 5.5 is a name without a shipped model behind it, and it is the name people are searching, so this is the page that answers for it. Read on 2026-10-12, the vendor's own model documentation lists exactly four current Claude models — Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 — and the Fable line's newest entry, Claude Fable 5.1, still carries the word "Latest" on its own page, dated September 1, 2026. Its predecessor Claude Fable 5 now sits beneath it under legacy models, with a migration note pointing upward. There is no row for a Claude Fable 5.5 anywhere in that set, no documentation path that resolves to one, and no price attached to the name. So the answer to "what is Claude Fable 5.5" comes in two halves, and this page gives both: what the name would mean in the Fable line if it arrived, and what you can actually call today, described in the present tense because today is when you are reading it.

Where a Fable 5.5 would sit, and where the line sits now

The Fable line is not an Opus variant. Anthropic's description of Claude Fable 5.1 calls it a Mythos-class model — a capability tier above the Opus class — positioned for demanding reasoning and long-horizon agentic work, and its own model page rates its comparative latency as the slowest of the four current models. The line's own internal naming is the only place a "5.5" has any meaning: Fable 5 shipped, Fable 5.1 followed it, and a Fable 5.5 would by that convention be the next step in the same series rather than a new tier.

Set beside the rest of the current lineup, the shape of that family is what a reader is usually trying to work out. Every figure below is the vendor's published rate, read from its models overview and pricing page on 2026-10-12, and every one of these four models carries the same 1M-token context window and the same 128K-token maximum output.

• Claude Fable 5.1 — $10 per million input tokens and $50 per million output tokens, described by the vendor as the choice for demanding reasoning and long-horizon agentic work, or when evaluations on Claude Opus 5.5 at higher effort still fall short. Comparative latency: slower. Default thinking effort: high, with adaptive thinking always on.

• Claude Opus 5.5 — $4 and $20 per million, positioned for long-running agentic coding and knowledge work. Comparative latency: moderate. Default effort: medium.

• Claude Sonnet 5.5 — $2 and $10 per million, the vendor's "best combination of speed and intelligence". Comparative latency: fast. Default effort: high.

• Claude Haiku 5.5 — from $0.10 and $0.50 per million, aimed at high-volume, latency-sensitive work such as classification, extraction and routing. Comparative latency: fastest. Default effort: medium.

A Fable 5.5 would inherit the first row's position by default and the first row's price bracket by convention — the vendor has kept Fable at $10 and $50 across both releases so far, and describes Fable 5.1 as extending Fable 5 "at the same input and output prices". That is a pattern, not a promise, and it is the reason the pricing section further down states plainly what is published and what is not.

What Anthropic documents for the shipped Fable, and what it does not

The capabilities that a reader would want attributed to a Fable model are documented — for Claude Fable 5.1, on its own pages, and not for anything called 5.5. The vendor's "what's new" page states that Fable 5.1 improves on Fable 5 and that the gap widens at higher effort levels, concentrating in six named areas: agentic coding across long sessions including multi-file features, large refactors and debugging; knowledge work producing finished documents, live-formula spreadsheets and slide decks; research and search on multistep and deep-research tasks; vision on dense charts, filings and tables nested in PDFs; reasoning across the full 1M-token context window; and computer use, including recovering from failed steps. Multilingual performance, the same page says, is on par with Fable 5.

What the vendor publishes no figure for is equally worth stating. There is no first-party benchmark table on the Fable 5.1 model page — no scores, no head-to-head, no evaluation result — and nothing resembling one for a Fable 5.5. Independent measurement exists for the shipped model only: Artificial Analysis, read on 2026-10-12 on its Intelligence Index revision v4.3.2, places Claude Fable 5.1 at 53, below Claude Sonnet 5.5 and Claude Opus 5.5 at 56 and an unnamed model at 58. Those are that organisation's measurements of a shipped model, not Anthropic's claims, and there is no matched-harness comparison between a Fable 5.5 and anything, because there is no Fable 5.5 to run.

The constraints are documented in more detail than the capabilities, and a few of them will surprise anyone arriving from a general model page.

• Forced tool use is rejected. Setting a tool choice of "any" or a specific named tool returns a 400 error on Claude Fable 5.1, on the stated grounds that a forced call would skip the always-on thinking pass. Schema enforcement moves to strict tool use with an automatic choice, or to structured outputs.

• Thinking blocks are bound to the model that wrote them. Only Claude Fable 5.1 and Claude Mythos 5.1 can read Claude Fable 5.1 thinking blocks, and editing anything earlier in a conversation invalidates the blocks after it — new accounts are checked from August 31, 2026, older ones only when the request opts in.

• Several sampling parameters are refused. Prefilling an assistant response returns a 400, and non-default temperature, top_p or top_k values return a 400. The minimum cacheable prompt is 512 tokens.

• Refusals are a first-class response. A declined request returns HTTP 200 with a refusal stop reason and a details object naming the policy area, and a refusal that arrives before any output may still be billed. The permitted fallback targets for Claude Fable 5.1 are Claude Opus 4.8 and Claude Opus 5.

• Data retention is 30 days, and zero data retention is not available for Claude Fable 5.1 unless Anthropic expressly authorises it, which the vendor states on the same page as the capability gains.

• Output carries provenance marks. Text from Claude Fable 5.1 carries Anthropic's statistical text watermark on every platform, and supported files produced through code execution carry signed content credentials when retrieved through the files API.

Context window and maximum output

For Claude Fable 5.1 the vendor publishes a 1M-token context window and a 128K-token maximum output, both stated on the model page and repeated in the four-column comparison on the models overview. Those two numbers are not a Fable-only figure: the same page gives 1M and 128K to Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 as well, so context length does not distinguish the tier — price and latency do. The Fable 5.1 page also carries a reliable knowledge cutoff of June 2026 and a training data cutoff of June 2026, identical to the other three.

For Claude Fable 5.5 there is no published context window and no published output ceiling, because there is no model page to publish them on. If you have seen a number for either one attributed to that name, it did not come from the vendor, and this page will not repeat it.

What goes in, and what comes out

This is the item most often guessed at, and the vendor answers it directly rather than by implication. The Claude Fable 5.1 model page states its input-to-output direction in one line: text and images in, text out. The models overview generalises it for the whole current lineup — every current model, it says, supports text and image input, text output, multilingual capability, vision and tool use. So a Fable 5.5 would be expected to accept images, if the lineup's stated baseline holds, but the honest form of that sentence is conditional: no vendor material describes a Fable 5.5's modalities at all.

Two details sit under the headline direction. Vision is not decorative for this line — the "what's new" page is explicit that the documented gains include reading dense charts, filings and tables nested inside PDFs, with crop-and-zoom tools on charts, which is why PDF-shaped input shows up in practice alongside images. And the output side is text only: nothing in the vendor's material for Claude Fable 5.1 describes image, audio or video generation from the model itself, though the code execution tool can produce supported media files as a side effect, which is what the provenance note above is about.

The rate card, and exactly what it covers

Anthropic publishes a full rate card for Claude Fable 5.1, and it is more granular than a headline pair of numbers. Read from the pricing page on 2026-10-12, in USD per million tokens: $10 input, $50 output, $12.50 for a five-minute cache write, $20 for a one-hour cache write, and $0.25 for a cache read — a quarter of the $1.00 cache read the vendor lists for Claude Fable 5, which is the reduction its own Fable 5.1 overview refers to when it says the new model extends the old one at the same input and output prices with cheaper cache reads. Batch processing takes 50% off both input and output.

What that price buys is the shipped model as configured: adaptive thinking that is always on, a default effort level of high, the 1M-token context and 128K-token output ceiling described above, vision and tool use, and the platform list on the model page. What it does not buy, and the vendor says so on the model's own page, is zero data retention.

There is no published price for Claude Fable 5.5. Not a preview price, not a range, not a placeholder on the pricing page — the Fable rows there are Fable 5.1, Fable 5 and Mythos 5.1, and the list stops. Any figure circulating for a Fable 5.5 is either a guess or a transcription of one of those three rows under the wrong name, and the reason to say so plainly is that a price is the single number people act on.

Where you can call the Fable line today

Anthropic lists five platforms for Claude Fable 5.1: the Claude API under the identifier claude-fable-5-1, Amazon Bedrock as anthropic.claude-fable-5-1, Google Cloud, Microsoft Foundry and Claude Platform on AWS — all five carrying the same identifier on the model page. That is the access surface for the shipped Fable, and it is the surface that has a row, a rate card and a retirement note saying not sooner than September 1, 2027.

Claude Fable 5.5 has no entry point on any of them. Requesting anthropic/claude-fable-5.5 from the OrcaRouter models API on 2026-10-12 returns model_not_found, and the same request for anthropic/claude-fable-5.1 returns a live record — a 1,000,000-token context window, a 128,000-token maximum completion, text, image and file inputs with text output, and a $10 / $50 rate. To be exact about what that means: we do not serve a Claude Fable 5.5, because there is nothing to serve. What we do serve is the Fable line as it exists — Claude Fable 5.1 and Claude Fable 5, alongside Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 — behind one OpenAI-compatible endpoint, one API for 200+ models, at 0% markup with the provider's list price passed through, so a change on Anthropic's own rate card reaches your bill the same day it reaches theirs.

What would have to change before this page changes

The useful thing about a reference page for a name with no model is that it can be checked, and it becomes wrong at a specific, visible moment rather than gradually. A Claude Fable 5.5 that shipped would produce four things at once, in this order.

• A row in the models overview. Four current names become five, and "Latest" moves off Claude Fable 5.1's page. Until that happens, the page that calls itself latest is the shipped one.

• A documentation path that resolves. Today the only Fable paths on the vendor's documentation sitemap are fable-5, fable-5/overview and fable-5-1 with its overview, migration guide and what's-new children. Nothing resolves under fable-5-5.

• A rate card with real numbers in it, including the cache write and cache read rows, because those are the numbers a migration decision actually turns on.

• A live identifier. The moment anthropic/claude-fable-5.5 stops returning model_not_found and starts returning a context window and a price, the question this page exists to answer has been answered by something better than this page.

Until then the practical advice is the same advice that applies to any unannounced string, and it is not exotic. Pin production to a versioned identifier and treat it as the contract — the vendor documents that a dateless identifier like claude-fable-5-1 maps to one fixed snapshot rather than floating to newer weights, which means your route cannot be upgraded underneath you by someone else's release. And when an unproven string does appear, put it behind automatic failover rather than in the critical path: a route that stumbles hands the request to a provider you already trust instead of handing your users an error, which is what makes it affordable to be early on a model nobody has measured yet. Both of those are properties of the endpoint, not of the model, and both keep working the day the fifth row appears.

One more thing belongs in a reference page, because it is the question that follows the search. A Fable 5.5 arriving would not make Claude Fable 5.1 disappear, stop being priced at $10 and $50, or lose its documentation; the two would coexist the way Fable 5 and Fable 5.1 do now, with a migration note between them and a legacy section below. The reader who needs a Fable model today does not have to wait for a name to become real in order to use the one that is.