Developers testing early-access builds spotted two unreleased model identifiers, claude-marshmallow-eap and claude-melon-eap, according to reporting from Odaily. Testers say Marshmallow performs better overall than Melon, and one tester described Marshmallow's conversational quality as better than the current Opus 5, though neither model is said to reach the level of Claude Fable 5. Anthropic has not confirmed the models exist, described what they do, or said anything about release timing.
This is a leak, not an announcement, and it's worth being precise about the difference. Everything above comes from unnamed testers relaying impressions on social platforms. None of it is a benchmark, a model card, or a statement from Anthropic.
Why this particular leak has more credibility than most
The "-eap" suffix, short for early access preview, is the part with an actual track record. In early July, an identifier called Claude Honeycomb EAP briefly appeared in the coding tool Cursor before being pulled, and Anthropic released Opus 5 sixteen days later, though the company never confirmed any link between the two. That pattern, an EAP codename surfacing in the wild followed by a real release weeks later, is why developers are treating Marshmallow and Melon as signal rather than noise. It's also why the community is guessing Marshmallow maps to a future Opus update and Melon to a Sonnet update, a guess Anthropic has done nothing to confirm or deny.
The timing adds context. Opus 5 has been out roughly a month, and Anthropic's own product leadership has already acknowledged in public feedback that the model can feel inconsistent, using the word "spiky" to describe it. A leak suggesting an internally-tested successor lands differently against that backdrop than it would in isolation.
Codenamed builds
Nothing, yet, and that's the discipline worth practicing. Codenamed builds get cancelled, renamed, or merged into products that look nothing like the leak suggested. The instinct to write "Anthropic's next Opus is coming" content off a Discord screenshot is exactly the kind of unverified claim that erodes trust with an audience that's increasingly good at spotting it. The better move is watching for an actual model card or pricing page, then reacting with substance instead of speculation.
That discipline, waiting for verifiable signal instead of chasing every leak, is the difference between a growth strategy that compounds credibility and one that burns it. If you want a second set of eyes on which AI stories are worth your team's attention this week, that's what our AI marketing services team is for.
Writing Team
.png)