Claude Fable 5.1 & Opus 5.1: Delay Extends as a Bedrock 404 Hints at Staging

Claude Fable 5.1 and Claude Opus 5.1 were expected on August 28, then rumored for August 31 and September 5. As of September 1 neither has shipped, and the newest signal is a staging clue rather than a launch: a community probe of Amazon Bedrock on August 31 found that a Claude Fable 5.1 identifier returns a distinctive 404 "model not found", while made-up identifiers return the ordinary 400 "invalid identifier" error. Nothing here is confirmed: Anthropic has announced neither model, no public API model ID exists, and this is a what-we-know-so-far piece. Every claim past the historical record is a leak, labeled as such.

Here is the short version. Claude Fable 5.1 and Claude Opus 5.1 are unannounced point releases of Anthropic's two top-tier models, Claude Fable 5 and Claude Opus 5. Community reports put Fable 5.1 in a grayscale rollout on Claude Web and link both models to the leaked "Melon" and "Marshmallow" early-access IDs. The release expected on August 28 reportedly slipped to "next week"; the August 31 date the trackers attached to Claude Fable 5.1 came and went without an announcement, and the Bedrock probe is now the only new datapoint. Whether the pair land together or separately — or at all — is unverified.

The signals: a delay that keeps extending, then a staging clue

The original signal is a single X post from August 29. @kimmonismus writes that "the assumption was actually that Fable 5.1 and Opus 5.1 were supposed to be released yesterday," that the release "seems to be delayed until next week," and that "some people already have access to the new models." It is one tweet from a community account — not a statement from Anthropic — and it relays an earlier "today is release day" reading from @synthwavedd. Treat the delay as a rumor about a rumor.

Two independent datapoints pointed the same way. WinCentral, citing leaker "Leo," reported on August 27 that Claude Fable 5.1 "is reportedly inside its launch window" — with the explicit caveat that the day in question was "a launch-window day rather than a guaranteed release day" and that the launch "could slip into early next week." And the DataLearnerAI release tracker listed rumored dates of 2026-08-31 for Claude Fable 5.1 and 2026-09-05 for Claude Opus 5.1 — both "next week," both unconfirmed. That "next week" has now partly elapsed: August 31 passed with no announcement, and the September 5 date for Opus 5.1 remains a tracker guess.

A "delay" is a strange thing to talk about for a model Anthropic has never announced, so it is worth being precise: there was never a public date for anyone to miss. What the community is tracking is a widely anticipated release window collapsing to something later than the grapevine expected. If the pattern holds — Claude Fable 5's own launch moved from a gray test to a near same-day full rollout — the slip is measured in days, not weeks.

The new datapoint: a Bedrock 404 that says "staged, not live"

On August 31, researcher @chetaslua — the same account that surfaced the Marshmallow and Melon early-access IDs in August — reported probing Amazon Bedrock for the anticipated identifiers, and the responses split in a way the community reads as meaningful. A Claude Fable 5.1 identifier returned 404 "model not found", while both a Claude Opus 5.1 identifier and a deliberately nonsensical one ("banana") returned 400 "the provided model identifier is invalid". The finding was picked up the same day by tech-news outlet RuntimeWire, which credits it to leo (@synthwavedd). This is a community probe, not a vendor statement, and nothing about it has been independently verified by Anthropic or by anyone with access to the AWS backend.

Why the split matters: on Bedrock, an incorrect model identifier normally produces a 400 validation error — the endpoint is telling you the string you sent is not a valid identifier at all. A 404 is the different case. It is what AWS returns when the requested identifier is recognized as a real model shape but the model is not available to that account, endpoint, or region. The working reading, in @chetaslua's words, is that Fable 5.1 is "identifiable but not staged": the name exists on the backend, and the availability switch is off.

Read the caveats before you read anything into it. A 404 can also appear when a valid model is unavailable in your region or denied to your account, and Bedrock responses vary by region, permissions, and inference-profile configuration. The probe says nothing about pricing, performance, context length, or a release date, and the public Bedrock catalog still lists Claude Fable 5 as the active model. What it adds is narrow but real: it is the first datapoint that separates the two .1 models. Fable 5.1 looks recognizable on AWS's backend; Opus 5.1 returned the same 400 as a nonsense string, which is the footprint of a name that does not exist there yet.

What the leak chain actually rests on

The Claude Fable 5.1 and Claude Opus 5.1 story has been building since July, and it helps to separate the layers, because they do not all carry the same weight:

• July 26 — an X thread from Pankaj Kumar claimed "Fable 5.1" was in preparation with a target August 2026 launch and unchanged pricing; re-shared the same day by Lumina and picked up by Times Of AI, which added that the model was "reportedly production-ready internally for some time," with a focus on long-horizon reasoning and agent work. Unverified.

• July 27 — Andrew Curran theorized that Fable 5.1 was "ready, but Anthropic are saving it for OpenAI's next release." Explicitly speculative, and the theory that now has a testable prediction: the release keeps slipping against a competitor's calendar.

• August 21–24 — researchers @chetaslua and @Lentils80 surfaced the early-access IDs claude-marshmallow-eap and claude-melon-eap, plus a reportedly live hidden-test variant claude-marshmallow-ht-ea…, appearing in third-party clients. The community reading maps Marshmallow to the Opus tier — a possible Claude Opus 5.1 — and Melon closer to Sonnet, but that mapping is contested and entirely unverified.

• August 25–27 — community tracker TestingCatalog flagged that Claude Web users are being routed from Claude Fable 5 to a newer checkpoint in the background, with an updated knowledge cutoff — the so-called grayscale rollout.

• August 31 — the Bedrock probe. @chetaslua reported that a Claude Fable 5.1 identifier returns 404 "model not found" on Bedrock, while a Claude Opus 5.1 identifier and a nonsense string ("banana") return the standard 400 "invalid identifier". Community reading: Fable 5.1 is recognizable on the backend but not live; Opus 5.1 has no such footprint. Unverified, region-dependent, and covered as news by RuntimeWire.

The confirmed surface, then, is still nearly empty. There is no model card, no pricing page, no comment from Anthropic, and no public API model ID — though the Bedrock probe is the first evidence that a Claude Fable 5.1 identifier exists somewhere on AWS's backend, returning a distinctive 404 rather than the standard invalid-identifier 400. The name "Claude Opus 5.1" has never been printed by any official source, and neither has "Claude Fable 5.1." What is real: the parent models exist, the EAP IDs appeared in third-party clients, some users' traffic is being routed to a newer Fable checkpoint, and a Fable 5.1 identifier appears to be recognized by Bedrock while an Opus 5.1 identifier is not. Everything else is inference.

The models a .1 would replace — and why the refresh matters

Any guess about what the new models change has to start from the models they would replace, because the current lineup already contains a pricing problem Anthropic created for itself.

• Claude Fable 5 — $10 / $50 per million tokens, 1M context, 128K max output, January 2026 knowledge cutoff, launched June 9 2026. Artificial Analysis Intelligence Index 62 (with fallback). Vendor-reported evals: SWE-bench Verified 95.5%, Terminal-Bench 2.1 88%, GPQA Diamond 94.1%.

• Claude Opus 5 — $5 / $25 per million tokens, 1M context, 128K max output, launched July 24 2026. Artificial Analysis Intelligence Index 63 at max and xhigh effort — the #1 model on that board. Vendor-reported evals: OSWorld 2.0 70.6% versus Claude Fable 5's 66.1%.

That is the tension underneath the leak: Claude Opus 5 undercuts the flagship at half the price while matching or beating it on several hard benchmarks. CursorBench 3.2 reportedly puts Claude Opus 5 at about 70% at $8.23 per task against Claude Fable 5's 70.5% at $17.32. When your cheaper sibling ties your flagship, the predictable move is a fast-follow refresh — which is exactly what a Claude Fable 5.1 would be. A separate Claude Opus 5.1 would push the value war one notch further: the whole top tier refreshes at once.

What a .1 refresh would plausibly fix is the complaint list, not just the benchmark list. Claude Fable 5's rollout drew steady community criticism about a 5-hour usage limit on the web app, aggressive safety classifiers that silently hand conversations to Claude Opus 5, heavy token consumption, and — per WinCentral's roundup of user frustration — laziness during longer tasks, excessive verbosity, and mid-task slowdowns. None of these fixes are confirmed in the leaks; they are the list anyone reading the room would bet on.

Why the window slipped — the timing theory

The leaker consensus has been that Anthropic is holding the refresh to counter OpenAI's next release, reportedly codenamed Astra. Andrew Curran's "saving it for OpenAI" theory was the early shape of that; later reports claim Anthropic decided not to wait and accelerated preparations instead. The delay to next week, if real, is consistent with a launch being gated on something other than readiness — and with a company that has iterated on a roughly two-week cadence all summer: Claude Fable 5 on June 9, Claude Sonnet 5 on June 30, Claude Opus 5 on July 24, and a 2M-context Sonnet 5.5 (codename "Fennec") leaked with a September window.

A model that is staged on a cloud backend but not serving fits the same pattern: the work exists, the availability switch is off, and the launch waits on rollout planning rather than readiness.

The gray-test precedent matters here. Claude Fable 5 itself launched by routing a small slice of users to the new checkpoint and sometimes rolling out fully the same day. That is why "some people already have access" is the least surprising part of the signal — it is how every recent Claude model has shipped. The delay is the unusual part.

"Already on it" — how people are checking

The community's working heuristic for detecting the new models is the knowledge cutoff. Claude Fable 5's official cutoff is January 2026, so a model that presents as Fable 5 but reliably answers questions about events after January is a candidate for being the grayscale successor. The same logic is being applied to the Opus side: a popular "anti-counterfeit" test prompt asks the model "opus 4.6 date of release and gpt image 1.5" without web search, on the theory that only a model trained past the original cutoff can answer it correctly.

Two cautions. A grayscale appearance or an early-access ID is not a release — Claude Opus 5 shipped July 24 after a "Honeycomb" EAP ID surfaced in a third-party client in early July and was never confirmed to be the same model. And the cutoff heuristic is a hint, not a proof: displayed cutoff metadata can change without a model swap, and routing varies by user and by surface. If you are building against a specific ID, wait for the ID to exist in a provider catalog.

What to do while the window slips

Nothing here is callable yet. Do not hard-code a guess at an anticipated model ID — the real string that ships may be different, and a leaked EAP handle is not an endpoint you should route production traffic through. Use the current, documented model IDs until an announcement lands. The Bedrock 404 changes none of that: an identifier that exists on the backend but returns "model not found" is still not an endpoint.

The useful work is the rehearsal. A first-week frontier release is the textbook case for automatic failover: when Claude Fable 5.1 or Claude Opus 5.1 is actually callable, the way to adopt it without betting a production path on it is to route a slice of traffic behind a fallback to a model you already trust. That is a routing rule, not a migration. On OrcaRouter, one API spans the whole current Claude lineup at provider list price with zero markup, so if the .1 models ship at the expected prices — $10/$50 for Fable 5.1, presumably $5/$25 for Opus 5.1 — that is the price you pay the same day, with no renegotiation and no second contract. (Those price expectations are themselves unverified.)

What to watch next

• The official model card and API ID — the only thing that converts any of this into a usable fact, and the exact string that ships is the test of which EAP sighting was real.

• The Bedrock catalog. The probe's 404 is a staging signal only if the identifier later appears as available — that is the confirmation test for the whole finding, and it would pin down the exact identifier string at the same time.

• The two rumored dates: August 31 for Claude Fable 5.1 has now passed with no announcement; September 5 for Claude Opus 5.1 remains on the tracker — both unconfirmed.

• Whether the pair land together or split. Reports vary between "Fable 5.1 and Opus 5.1" and "Fable 5.1, Sonnet 5.1, Opus 5.1" as a full product-line refresh.

• The price sheet. An unchanged Fable 5.1 at $10/$50 against a $5/$25 Opus 5.1 keeps Anthropic's internal price war running — the number to watch if you route traffic by cost.

• Anthropic's silence. The company does not comment on gray tests, so an update to a model card or the docs is often the first official confirmation.

The honest summary is a question. Did Claude Fable 5.1 and Claude Opus 5.1 slip to next week — or was August 28 never a real date in the first place? The leak chain says both models exist and some users already appear to be on them; the Bedrock probe adds that a Claude Fable 5.1 identifier is currently recognizable on AWS's backend while a Claude Opus 5.1 identifier is not. As of September 1, neither is callable through any public API. None of it is confirmed. Watch Anthropic's official channels; the silence is still the loudest part of the story.

FAQ

Is there a confirmed release date for Claude Fable 5.1 or Claude Opus 5.1?

No. Anthropic has announced neither model. The August 28 expectation and the August 31 date attached to Claude Fable 5.1 have both passed without a release; the only date still in circulation is a tracker's rumored September 5 for Claude Opus 5.1 — all unverified.

How can I tell if I have already been given access to Claude Fable 5.1 or Claude Opus 5.1?

The community heuristic is the knowledge cutoff: ask a model presenting as Claude Fable 5 about a well-known event after January 2026, its official cutoff, and see whether it answers correctly. There is also a popular "anti-counterfeit" prompt that asks for the Claude Opus 4.6 release date without web search. These are hints, not proofs, and neither confirms which model you are actually on.

Does grayscale or early-access access mean the models have been released?

No. Some users being routed to a newer checkpoint on Claude Web, or an EAP ID appearing in a third-party client, is testing — Claude Opus 5 shipped July 24 after a "Honeycomb" EAP ID appeared in early July and was never confirmed to be the same model. A release is an official announcement with a model card and a documented API ID.

Does the Bedrock 404 mean Claude Fable 5.1 is released or callable?

No — it means the opposite. A 404 "model not found" on Bedrock is what you get when an identifier is recognized but the model is not available to your account or region; a model that is actually serving would return a success response or a permissions error, not a not-found. Neither .1 model has a model card or a public API model ID, and Bedrock's catalog still lists Claude Fable 5. The probe is evidence of staging work, not a launch.