Anthropic appears to be developing two unreleased Claude models, codenamed Marshmallow and Melon, that surfaced in third-party developer apps and Discord discussions on August 24. The identifiers, labeled as early access preview (EAP) versions, suggest the company may be preparing updates across its entire model lineup. While Anthropic has not officially confirmed either model, the leak offers the first concrete clues about how the company plans to address recent criticism over Claude Opus 5's inconsistent performance.

What Are Marshmallow and Melon?

The two model identifiers are claude-marshmallow-eap and claude-melon-eap. Based on early tester feedback and community speculation, Marshmallow is widely believed to be Claude Opus 5.1, while Melon is thought to be either Claude Sonnet 5.1 or a new Haiku-class model. However, these identities remain unconfirmed by Anthropic.

Early reports from developers who accessed the models describe both as performing solidly, though neither appears to reach the level of Claude Fable 5, Anthropic's most advanced model. What stands out to testers is that Marshmallow reportedly performs better than Melon and may already outperform the current Claude Opus 5 on certain tasks. Both models are noted for their ability to generate 3D content in a single pass and for using substantial token counts during reasoning processes.

Why Does the Timing Matter?

The EAP label is significant because Anthropic has used it before to signal models approaching public release. The most recent precedent is Claude Honeycomb, which appeared in Cursor's model picker on July 8 and 9, 2026, before being removed. Approximately 16 days later, on July 24, Claude Opus 5 officially launched. This timeline suggests that Marshmallow and Melon could arrive within a similar window, though the exact launch date remains unknown.

The current Claude lineup spans multiple price and capability tiers. Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens. Claude Opus 5 is priced at $5 and $25 per million tokens respectively. Claude Sonnet 5, released June 30, 2026, costs $3 and $15 per million tokens. The most affordable option, Claude Haiku 4.5, launched October 15, 2025, and costs $1 and $5 per million tokens.

What Problem Are These Models Solving?

Anthropic has faced criticism since Opus 5's launch over what users describe as inconsistent, lazy, and verbose outputs. The company's own Thariq acknowledged that Opus 5 is "really spiky" and stated that Anthropic wants Claude to feel "more consistent and warm." This feedback directly shaped expectations around Marshmallow, with early testers reporting that it delivers solid results even on medium reasoning settings, suggesting more reliable performance without requiring the highest effort levels.

How to Track Upcoming Claude Model Releases

• Monitor Developer Platforms: Watch Cursor, third-party API integrations, and Discord communities where model identifiers often surface before official announcements, giving early signals about unreleased versions.
• Check Anthropic's Official Channels: Follow Anthropic's website, blog, and social media accounts for official model card releases, API listings, and pricing updates that confirm new versions.
• Review Model Pricing and Context Windows: Compare specifications like token costs, context window sizes, and maximum output lengths to understand where new models fit in the product lineup and their intended use cases.
• Track Release Patterns: Study the timeline between EAP leaks and public launches, which historically span two to three weeks, to estimate when new models might become available to the general public.

Some community posts suggest Anthropic may be accelerating updates across its top, mid, and budget tiers simultaneously, which would represent a faster release cadence than the company has maintained in the past. However, internal codenames do not always translate to public releases. Claude Numbat, for example, remained an internal identifier for months without launching publicly.

As of now, Anthropic has not confirmed the existence of Marshmallow, Melon, Opus 5.1, or Sonnet 5.1. The company continues to compete in a rapidly evolving AI landscape where Google is preparing Gemini 4, with some reports pointing to a possible December launch following delays to Gemini 3.5 Pro. Until Anthropic makes an official announcement, all details about these models remain speculation based on leaked identifiers and early tester feedback.