Claude Marshmallow & Melon EAP Leak: Two Unreleased Anthropic Models, One Already Live

On August 21, the model ID that researcher @chetaslua says was actually live carried a string Anthropic has never announced: claude-marshmallow-ht-ea…, truncated exactly there in the source's proof. Around it, two sibling early-access IDs surfaced in developer tools and Discord communities over the following days — claude-marshmallow-eap and claude-melon-eap. Anthropic has not acknowledged any of them: no model card, no documentation page, no API listing. "Not a single first party byte," in chetaslua's words — the same researcher who, in mid-August, decoded a hidden "Kettle" route identifier in the signed thinking of Claude Fable 5. This is a leak report: every concrete detail below is unverified until the vendor says otherwise, and nothing here should be read as a launch.

What the leak actually says

The sightings arrived in two X posts this week. @Lentils80 first flagged the pair of early-access IDs, and multiple outlets traced the report back to that post; @chetaslua followed with the sharper detail, and it is chetaslua's framing that makes this more than a name-drop. The "eap" suffix is Anthropic's early-access-program marker — the same label that preceded the "Honeycomb" codename in July.

• The two IDs are claude-marshmallow-eap and claude-melon-eap, spotted in third-party developer applications and relayed through Discord communities; the "eap" suffix marks them as early-access builds.

• Per chetaslua, neither ID has a single first-party byte — no Anthropic documentation, no model card, no official API entry. They exist only in third-party sightings.

• The source speculates the models may be restricted to red-team access only. That is speculation, not confirmation, and nothing official offers either ID.

• The ID that was actually live on August 21, per chetaslua, was claude-marshmallow-ht-ea… The "-ht-" segment is a plausible hidden-test marker, but that reading is interpretation, not fact, and the full string truncates in the source.

That last point is the one worth sitting on. An "-eap" ID sitting in a client's model picker is one thing; a variant the leaker says was live is another. If the screenshot shows what it claims, there is a Marshmallow-branded model with real server-side presence behind the "-ht-" label, even though nothing official calls it by name.

What the testers claim — and what that's worth

Roundups from Odaily, TokenPost, and others relay a consistent but entirely anecdotal picture from developer Discord communities:

• Marshmallow is reported to outperform Melon overall.

• One tester described Marshmallow's conversational experience as better than Claude Opus 5.

• Neither model is reported to reach Claude Fable 5's level.

No scores, no leaked evals, no benchmark outputs — the reporting is clear that these are tester impressions, not measurements, so treat them as flavor. The claims are coherent, though. Claude Opus 5 is positioned for agentic coding and enterprise work; a follow-on Opus-class model that improves conversational polish is exactly what a point release would do. And "not at Fable 5 level" matches the shape that produced Opus 5 in the first place — a model that wins the independent leaderboard while the flagship tier sits above it on price.

Where they'd slot in the Claude lineup — and the pricing math

Anthropic's current model line, at list price per 1M tokens, reads like this:

• Claude Fable 5 — $10 input / $50 output, 1M context, Intelligence Index 62 (with fallback).

• Claude Opus 5 — $5 / $25, 1M context, Intelligence Index 63 at max and xhigh effort — the #1 model on the Artificial Analysis board.

• Claude Sonnet 5 — $2 / $10, 1M context, Intelligence Index 55 at max.

• Claude Haiku 4.5 — $1 / $5, 200K context, Intelligence Index 30.

Community speculation maps Marshmallow to the Opus tier — a possible "Opus 5.1" — and Melon closer to Sonnet, with a minority reading that one of them could be a new Haiku. Nothing is confirmed, but the pricing math makes the Opus-tier reading the interesting one. Anthropic's own lineup already contains a value problem: Claude Opus 5 sits at the top of the independent Intelligence Index at $5/$25, exactly half the price of Claude Fable 5 at $10/$50, and CursorBench 3.2 shows why — Opus 5 scores about 70% at $8.23 per task against Fable 5's 70.5% at $17.32. SemiAnalysis-reported spend data has Fable 5 at roughly 11% of enterprise Claude usage. An Opus-class Marshmallow at Opus pricing would not need to beat Fable 5 to change a decision; it would only need to exist, because the price gap would do the arguing.

The cadence is the tell. Anthropic shipped Claude Fable 5 on June 9, Claude Sonnet 5 on June 30, and Claude Opus 5 on July 24. Roughly two weeks after Opus 5, the next Sonnet — internal codename "Fennec," a 2M-context Sonnet 5.5 aimed squarely at DeepSeek V4 Flash's cost-effectiveness — leaked with a September window. A week after that, two more food-coded EAP IDs. Anthropic is iterating on roughly a two-week cycle, and it is doing so under visible pressure: DeepSeek V4 Flash sits at Index 52 at a fraction of Claude's prices, and OpenAI's GPT-6/Astra cadence is the backdrop to every flagship decision this quarter.

The naming is a pattern — and so is the timing

Anthropic's public model names are literary genres: Haiku, Sonnet, Opus, Fable, Mythos. Its internal codenames were animals — Fennec for the upcoming Sonnet 5.5. The new batch is food: Honeycomb in July, Marshmallow and Melon now. The shift is not cosmetic trivia; it points at a development branch that has nothing to do with the literary product line, which is why the community treats these as new-build test candidates rather than simple renames.

The Honeycomb precedent is the one worth holding onto. In early July, claude-honeycomb-eap briefly appeared in the Cursor code editor and then vanished. Anthropic launched Claude Opus 5 on July 24 and never confirmed that Honeycomb was it. The lesson for this leak: an EAP codename surfacing in a third-party client has, this year, preceded a real launch by roughly three weeks. It is not proof — codenames can be renamed, merged, or cancelled — but it is the strongest available prior for reading the room.

What a developer should do right now

Nothing here is callable. claude-marshmallow-eap, claude-melon-eap, and the reportedly live claude-marshmallow-ht-ea… variant are not offered through Anthropic's public API, and the red-team-only reading is the most plausible explanation for the "live" signal. Do not build against a leaked ID.

The concrete thing to watch is the official model ID. A genuine release should appear in Anthropic's model documentation under a proper name — and because every Claude model ID is a pinned snapshot, the exact string that ships is itself a fact worth recording. When it appears in a provider catalog, that is confirmation, not rumor.

In the meantime, the current Claude lineup is what you can actually route: Claude Opus 5 at $5/$25, Claude Sonnet 5 at $2/$10, Claude Fable 5 at $10/$50. On OrcaRouter those are served at Anthropic's list price with zero markup, so a vendor price cut is live the same day — no renegotiation, no second contract. When and if a Marshmallow or Melon appears in a provider catalog, the same pass-through rule applies to it.

The unproven-model play is the one worth rehearsing now. The way to adopt a model you cannot verify is not to bet a production path on it — it is to route a slice of traffic to it with automatic failover. Point your hardest agentic tasks at Claude Opus 5 today with a failover rule to Claude Sonnet 5, and when the new models land the change is one routing-DSL line instead of a rewrite. One API for 200+ models, and the switch is a rule, not a migration.

What we're watching

• The official model ID. A new ID in Anthropic's docs or a provider catalog is the most concrete confirmation signal — and the exact string (claude-marshmallow-ht-…, claude-marshmallow-eap, or something else entirely) will tell us which of the sightings was real.

• Which tier. Marshmallow as "Opus 5.1" vs Melon as Sonnet-class — or a new Haiku that nobody is predicting.

• The price sheet. An Opus-class model at $5/$25 would pressure Claude Fable 5's $10/$50 premium from inside the family. That is the number to watch.

• Sonnet 5.5's September window. If Marshmallow or Melon lands before or beside Fennec, Anthropic's cadence is even faster than the leaks imply.

• Whether Anthropic ever confirms the Honeycomb → Marshmallow → Melon throughline, or the codenames quietly die.

This is a leak piece, so the honest summary is a question. Two food-coded Claude IDs with not a first-party byte between them, and a hidden-test variant that was apparently live on August 21 — that is the entire verified surface. It matters because the pattern has a track record this year: an EAP codename surfaced in July, and Claude Opus 5 shipped three weeks later. If that rhythm holds, an Opus-class Marshmallow and a Sonnet-class Melon are closer than Anthropic's silence suggests. Until an official ID appears in a model card, treat every capability claim as rumor, keep calling Claude Opus 5 or Claude Sonnet 5 for real work, and keep the option to route around the announcement when it comes.