Claude Marshmallow & Melon Leak: Is Opus 5.1 the Answer to Opus 5's Critics?

Claude Opus 5 has been shipping for exactly a month, and Anthro​pic's own team has already conceded the model is "really spiky." That admission is the backdrop for the leak that broke on August 24: the identifiers claude-marshmallow-eap and claude-melon-eap surfaced in third-party developer apps and Discord, and the community has read them as Claude Opus 5.1 and possibly a new Claude Haiku. None of it is announced — no model card, no release date, no comment from Anthro​pic. This is a what-we-know-so-far piece on a leak that looks like the vendor answering its critics, and everything in it beyond the historical record is unverified.

The suffix is the tell. "-eap" is Anthro​pic's early-access-preview marker, and the last time an identifier like it appeared in the wild, a real model followed within weeks. In early July, Claude Honeycomb EAP showed up in Cursor's model picker and was pulled within hours; sixteen days later Claude Opus 5 went live. If Claude Marshmallow and Claude Melon are on the same path, they are not vaporware — they are the next two entries in the Claude family, spotted a few beats before their announcement.

What actually leaked

Per the reports circulating on August 24, the raw material is thin and mostly names:

• Two new identifiers — claude-marshmallow-eap and claude-melon-eap — have been spotted in Anthro​pic's API surface and in Discord. One account spells the first "claude-mashmallow-eap," a one-letter typo that reads like a real string transcribed from somewhere rather than a cleanly fabricated list.

• @kimmonismus, who surfaced the pair, reads them as "probably an Opus update and maybe even a new Haiku."

• A second report, from the account J A Z I I, describes the models as "not fable class models but decent, so mostly opus or sonnet models" — agreeing on Opus, but offering a Sonnet rather than a Haiku reading for the smaller of the two.

• The release is described as "likely imminent," with no date attached.

That is close to the entire factual payload. Anthro​pic's documentation, newsroom, and price sheet mention neither name. Marshmallow and Melon are working codenames — and notably, they break the animal naming scheme Anthro​pic's internal builds have usually followed (Fennec, Numbat). Whether the switch to food names signals a new internal branch is speculation.

The last time an "-eap" identifier leaked, a real model followed

A leaked model name floating around X is cheap. An identifier ending in "-eap" is a different class of signal, because it implies a real build exists and is being staged for release. The Honeycomb episode is the closest thing to a controlled study of how Anthro​pic leaks actually behave:

• July 8–9, 2026 — Claude Honeycomb EAP appears in Cursor's model picker with a 1M-token context window, an "extra high" effort mode, and a description along the lines of "Anthro​pic research model with per-turn controls and safety fallbacks. Early access preview." It is pulled within hours.

• July 24 — the string claude-opus-5 appears in a Cursor error dialog, and Claude Opus 5 officially launches the same day.

So the distance from "EAP identifier visible in a shipping product" to "GA launch" was about two and a half weeks for Opus 5. If claude-marshmallow-eap and claude-melon-eap surfaced from the same kind of staging, a late-August or early-September release is a reasonable expectation. But that cadence is one data point, not a law — the unreleased Numbat codename sat quietly for months, a reminder that identifiers can also surface and go nowhere.

The signal behind the signal: Anthro​pic is answering its critics

What makes this leak more than a catalog sighting is what the community was already telling Anthro​pic about Claude Opus 5. The recurring complaint, voiced bluntly by @kimmonismus and echoed across the timeline, is that Opus 5 is "lazy, sloppy, and verbose" — inconsistent in personality and effort depending on the setup, rather than the "consistent and warm" model Claude users are used to. Kim put it as a truism: "Opus 5 isn't good, and they know it."

Anthro​pic's response has been unusually public. Thariq, on the Claude team, agreed in reply that "Opus 5 is a really spiky model," said the company wants models to be "consistent and warm and feel like Claude," and called improvement the highest priority. Another Anthro​pic voice, Boris, suggested people were "sleeping on using Opus to hill climb" for optimization tasks — a point Kim found evasive, the kind of "you're using it wrong" framing that was itself part of the problem.

The leak ties into that thread directly. In the same week, the reports on the new models claim they "show quite good results, even on Medium reasoning," with "(Opus 5.1?)" floated as the likely identity — and the commenter reads the timing as evidence "Thariq and Boris reiterated yesterday that they had taken the criticism to heart." If the EAP builds are what the team was polishing while the criticism peaked, the "good even on Medium reasoning" observation is exactly the kind of detail that would mark a fix: a model whose quality does not fall off a cliff when you dial effort down from the maximum. That remains a leak claim, not a benchmark — but it is a plausible response to the specific complaint.

What Marshmallow and Melon probably are

To weigh the two readings, line up the current Claude family and where each tier last moved:

• Claude Fable 5 — $10 in / $50 out per 1M tokens, 1M context, 128K max output. GA June 9, 2026. The flagship, and the most expensive.

• Claude Opus 5 — $5 / $25 per 1M, 1M context, 128K max output. GA July 24, 2026. Thinking on by default, with several effort levels.

• Claude Sonnet 5 — $3 / $15 per 1M (intro $2 / $10 through August 31), 1M context. GA June 30, 2026. Near-Opus quality on coding and agentic work at roughly 60% of the price.

• Claude Haiku 4.5 — $1 / $5 per 1M, 200K context. GA October 15, 2025 — the oldest model in the lineup by about ten months.

Against that board, the "Opus update" reading of Marshmallow is easy to credit. Anthro​pic's Opus cadence in 2026 has been unusually fast — Opus 4.6, 4.7, and 4.8 shipped across the first half of the year, and Opus 5 followed Opus 4.8 by only eight weeks. A point release at the same $5 / $25 price, fixing the behavior complaints rather than pushing raw capability, is well within pattern, and an EAP staging build is exactly the shape such an update would take.

Melon is where the two sources split, and it is the more interesting question. A new Haiku is overdue on age alone: Claude Haiku 4.5 has been the workhorse tier for ten months while everything above it moved twice. A Haiku-class successor at $1 / $5 — or an introductory cut — would matter enormously for high-volume traffic, classification, summarization, and anything where token cost is the binding constraint. The alternative reading, Sonnet-class, would be a mid-cycle refresh rather than a new tier: an incremental step under the current Claude Sonnet 5, closer in spirit to what a 4.6 was. The honest position is that the pair most likely means "one Opus-family model and one smaller model," with the smaller model's tier unconfirmed.

Why the release is coming now — and why it feels rushed

The timing reports emphasize one word: urgent. This is described as a faster turnaround than Anthro​pic's usual cadence, arriving roughly two months after the Fable 5 launch — and there is a commercial reason for the hurry. Claude Fable 5 has had a difficult first two months: it shipped with stealth guardrails that quietly reduced response quality on cutting-edge topics (a move aimed at deterring distillation), a mandatory 30-day data-retention policy that complicated enterprise adoption, and media-reported enterprise uptake of only around 11% of Anthro​pic's spend share versus roughly 23–25% for Ope​nAI's GPT-5.6 Sol. Meanwhile open-source models have gone from about 11% to 62% of token share in four months, per the same reporting, putting pressure on every proprietary price point.

Claude Opus 5 was the first step of the answer — cheaper than the flagship and competitive with it on cost-adjusted coding work (on CursorBench, Opus 5 scored 70% at about $8.23 per task versus Fable 5's 70.5% at about $17.32). Marshmallow and Melon look like the follow-through: a fixed, smoother Opus at the same price, plus a cheaper small model to defend the high-volume tier that open-source is attacking. That is a coherent product strategy, but the leaks being "likely imminent" suggests the models are being rushed out to meet a competitive window rather than after a leisurely development cycle — which cuts both ways for buyers. First-week releases are exactly when a point release can be uneven.

What to do while you wait

The boring answer first: neither Claude Marshmallow nor Claude Melon exists as a callable model today. There is no API ID, no provider catalog entry, and nothing to build against — any "early access" listing you can pay for right now is not what this leak describes. The claude-opus-5 and claude-haiku-4-5 identifiers you are already calling are unchanged, and until Anthro​pic announces new ones, keep calling what you are calling.

The useful work is to make the release cheap to adopt when it lands. New Claude models typically appear in provider catalogs within hours or days of announcement, and that is where a routing layer earns its keep: on OrcaRouter, one API spans 200+ models, and provider list price is passed through with zero markup — so when a new Opus or Haiku actually appears in a catalog, it is callable through the same key and endpoint at the vendor's own price, live the same day, with no second contract and no renegotiation. An unproven first-week model is also the textbook case for automatic failover: pin the new model with a fallback to Claude Opus 5, or give it a shadow share of traffic, so a half-baked launch becomes a routing rule instead of an incident. That posture costs nothing today and turns a leak into a head start when the announcement lands.

What we're watching

• Whether Anthro​pic announces the models. The Honeycomb precedent says an EAP sighting can precede GA by two and a half weeks — or the identifiers can sit quietly for months, as the unreleased Numbat codename did.

• The shipping names. Marshmallow and Melon are working codenames; the announced names are the strategic signal — Opus 5.1 vs Opus 5.5 vs Opus 6 says different things about Anthro​pic's cadence.

• The API model IDs. A genuine new version ships under a new ID; the exact strings that appear in a provider catalog are the single most concrete thing to watch.

• Pricing. Whether Marshmallow holds Claude Opus 5's $5 / $25, and whether a new Haiku keeps $1 / $5 — the Haiku price is the bigger tell, because it reveals whether Anthro​pic intends to defend the budget tier against open-source.

• The Haiku-vs-Sonnet question. The two leak readings disagree about the smaller model; whichever tier Melon actually fills changes what the release means for high-volume workloads.

Frequently asked questions

Is Claude Marshmallow out yet?

No. The identifiers claude-marshmallow-eap and claude-melon-eap have been spotted, but nothing has been announced and Anthro​pic has not commented. Until there is an official model card or a catalog entry, treat every detail — including the "Opus 5.1" reading — as unverified.

What does "-eap" mean on a model identifier?

It stands for early access preview, and it is the same suffix that preceded Claude Opus 5: the Claude Honeycomb EAP build appeared in Cursor in early July and Claude Opus 5 launched sixteen days later. An EAP identifier implies a real staging build exists, which is stronger evidence than a bare rumor — but it is not a guarantee.

When will Anthro​pic release the new Claude models?

No date is confirmed. The leak says a release is "likely imminent," and the Honeycomb precedent suggests EAP sightings can precede GA by roughly two weeks, but that cadence is a single data point — the unreleased Numbat codename shows identifiers can also sit for months. Watch for an Anthro​pic announcement, not the leak thread.

Is Claude Opus 5.1 real?

Unverified. The community reads Claude Marshmallow as an Opus-class update — the working "Opus 5.1" label — and early tester reports claim its conversational behavior is smoother than Claude Opus 5's. None of that is confirmed, and the shipping name may be different from the codename, so treat "Opus 5.1" as a hypothesis, not a product.

Will the new models be available through OrcaRouter?

When and if Claude Marshmallow and Claude Melon appear in a provider catalog as shipping models, the same rules apply as for every other model we route: one API, provider list price passed through with zero markup, automatic failover, and no second contract. Until then there is nothing to call — the models do not exist as endpoints yet.

This is a leak piece, so the honest summary is a question. Two early-access identifiers have surfaced that look like the next Opus-family model and the next small model, they have landed in the middle of Anthro​pic's most public reckoning with Claude Opus 5's behavior, and the last time an EAP identifier surfaced this way, a real launch followed within weeks. Nothing is verified until Anthro​pic says it, and the things worth watching are concrete: the announcement, the shipping names, the model IDs, and — if a new Haiku is really coming — the price. If you are calling Claude Opus 5 or Claude Haiku 4.5 today, keep calling them, and keep the option to route around a rumor open, which is exactly what a failover rule is for.