RUMOR: two candy-named test strings hint at the next Claude

Editorial label: RUMOR. While OpenAI and xAI trade blows over GPT-6 Astra and the promised Grok 4.7, Anthropic has its own trail of crumbs. Between August 21 and 24, 2026, two early-access identifiers, claude-marshmallow-eap and claude-melon-eap, surfaced in third-party developer apps and Discord communities. The "-eap" suffix is the same naming shape Anthropic used for claude-honeycomb-eap, a build that appeared in Cursor's model picker in early July and turned into Claude Opus 5 sixteen days later. Anthropic has not confirmed either string, and no official page mentions them.

Why these two names, and why the guessing got specific

Developers who spotted the identifiers describe Marshmallow performing noticeably better than Melon in informal tests, and stronger than the current Claude Opus 5 on some tasks, though neither reportedly reaches Fable 5 territory. That gap led most observers to read Marshmallow as an Opus-tier successor, tentatively called Opus 5.1, with Melon mapped to Sonnet 5.1. A minority reading argues Melon could instead be an overdue new Haiku, since Haiku 4.5 has held the entry tier for ten months while every other rung moved twice.

The timing lines up with Anthropic's own admission of a rough edge

The leak lands in the middle of unusually public criticism of Claude Opus 5, which shipped July 24. Users have complained for weeks about verbosity, jargon and inconsistent quality outside maximum reasoning effort, and Anthropic's own team member Thariq publicly agreed Opus 5 is "really spiky," saying the company wants its models to feel more "consistent and warm." If Marshmallow and Melon are what the team has been polishing in response, a point release fixing exactly those complaints would be a logical next step, not a leap in raw capability.

A second, separate signal raises the stakes: Model 2

Ten days before the leak, Anthropic's own August 14 risk report disclosed, for the first time, an internal system called Model 2, described as a noticeable improvement over the company's flagship Mythos 5 on internal coding and research benchmarks, with no plans announced to release it externally. Nothing ties Model 2 directly to Marshmallow or Melon, and Anthropic has not linked the two threads. But the clustering of an internal capability disclosure with a pair of leaked early-access strings, right as the company also prepares a $2 trillion IPO, is the kind of pattern that has preceded real Anthropic launches before.

What would confirm or kill the rumor

The Honeycomb precedent suggests roughly two and a half weeks between an EAP sighting and a public launch, which would point to early September. But Anthropic has also let codenames sit unused for months without ever shipping them, so the clock is not a rule. Confirmation would look like an official model card, a live API identifier containing "opus" or "sonnet" plus a version number, and published pricing. Until one of those appears, "Opus 5.1" and "Sonnet 5.1" remain inferred labels, not Anthropic's own naming.

What changes in MaxAssistant's read

This is a rumor with a real artifact behind it, not an anonymous claim: the strings were seen circulating by multiple independent developers, and the pattern matches Anthropic's own established leak behavior. That is enough to justify getting ahead of the story, but not enough to call it a launch. Anyone building on Claude Opus 5 today should keep watching for the identifier drop rather than planning around Opus 5.1 as if it already existed.

Sources

CellCog, Claude Opus 5.1 and Sonnet 5.1: Release Date, Leaks, and What We Actually Know: https://cellcog.ai/blog/claude-opus-5-1-release-date/ | OrcaRouter, Claude Marshmallow & Melon Leak: Opus 5.1 and a New Haiku?: https://www.orcarouter.ai/blog/claude-marshmallow-melon-leak | Tech Insider, Two Claude Models Leak as Anthropic Eyes $100B IPO: https://tech-insider.org/two-claude-models-leak-anthropic-100b-ipo-2026/