Anthropic Claude Marshmallow and Melon Models Leaked

Anthropic Claude Marshmallow and Melon Models Leaked

Two previously unannounced Claude model codenames have surfaced in Anthropic’s API, giving developers an early look at what could be the company’s next generation of language models.

The leaked identifiers, claude-marshmallow-eap and claude-melon-eap, were reportedly discovered through API access and subsequently confirmed by multiple developers across the developer community and Discord.

The eap suffix is widely interpreted as Early Access Program, suggesting that both models are currently undergoing limited testing rather than representing finalized public releases.

The discovery comes only a short time after reports of Anthropic’s leaked Sonnet 5.5 model, reportedly associated with the internal codename Fennec. If the latest sightings are genuine, Anthropic appears to be maintaining an increasingly aggressive model development and testing cycle.

🚨 Two New Claude Model Codenames Surface

#

The two newly exposed identifiers are:

claude-marshmallow-eap
claude-melon-eap

Both names appeared in Anthropic’s API environment before any corresponding public announcement.

Community reports indicate that developers were able to identify the codenames independently through API access, with subsequent discussion spreading across developer communities and Discord.

The eap suffix is particularly significant because it likely indicates an Early Access Program build. This would place the models somewhere between internal development and general availability.

At this stage, there is no reliable indication that either codename represents a finalized product name.

What the leaked names reveal

#

The codenames themselves do not identify the models’ eventual product tiers. Current community speculation includes:

• A future Opus-generation model.
• A new Sonnet-generation model.
• An updated Haiku model.
• Experimental models belonging to a separate development branch.

Neither Marshmallow nor Melon appears to correspond directly to Anthropic’s established public model naming tiers.

🧪 Early Access Status and Developer Reports

#

The strongest technical clue is the eap suffix.

If eap indeed represents Early Access Program, the models are likely being exposed to a limited developer population for evaluation, benchmarking, and feedback before broader deployment.

This distinction is important because API visibility does not necessarily indicate that a model is production-ready. Early-access models can change substantially before release, including modifications to:

• Model architecture.
• Inference behavior.
• Context handling.
• Tool-use capabilities.
• Safety policies.
• Latency and throughput characteristics.
• Pricing and availability.
• Final model naming.

Consequently, the current codenames should be treated as development identifiers rather than confirmed commercial products.

Marshmallow reportedly leads Melon

#

Early tester reports reportedly favor Marshmallow over Melon, with some developers claiming that Marshmallow delivers a stronger overall conversational experience.

More notably, community feedback has suggested that Marshmallow may feel better in conversation than the current Opus 5.

These reports remain anecdotal, however. Without controlled benchmarks or official evaluation data from Anthropic, they should not be interpreted as evidence that Marshmallow objectively outperforms Opus 5 across coding, reasoning, tool use, or other workloads.

🏷️ Anthropic Appears to Be Changing Internal Codename Themes

#

The newly exposed names are also interesting because they appear to depart from Anthropic’s previously reported internal naming conventions.

Anthropic’s public model families have historically used literary terms such as:

• Haiku
• Sonnet
• Opus
• Fable
• Mythos

Internal development codenames reportedly followed a different pattern, including animal-based identifiers.

For example, the previously reported internal codename for Sonnet 5.5 was Fennec, referring to the desert fox.

The latest names, however, are distinctly food-oriented:

Marshmallow
Melon

This could indicate a new internal naming convention or a separate development branch.

Why internal codenames matter

#

Internal codenames rarely provide enough information to determine a model’s final capabilities or product positioning. However, consistent changes in naming patterns can sometimes provide clues about how an organization separates development projects internally.

If Marshmallow and Melon belong to a new model family, the naming shift could reflect a new generation of internal experimentation rather than simple iterations of existing Opus or Sonnet models.

That interpretation remains speculative until Anthropic provides additional information.

⚡ Anthropic’s Model Release Cadence Is Accelerating

#

The timing of the leak is potentially more significant than the names themselves.

The reported discovery follows closely after the leak surrounding Sonnet 5.5. Anthropic has also been moving through model iterations at a rapid pace, with major updates reportedly appearing at intervals of only a few weeks.

For context:

• Sonnet 5 was reportedly launched on June 30.
• Opus 5 was released on July 24.
• The latest Marshmallow and Melon sightings indicate that additional models may already be entering limited-access testing.

If the EAP models progress successfully, a public announcement could follow relatively soon.

However, API exposure should not be treated as a reliable release schedule.

EAP does not guarantee a public launch

#

An early-access model can be:

• Renamed before release.
• Folded into another model.
• Replaced by a newer checkpoint.
• Restricted to selected developers.
• Delayed indefinitely.
• Canceled entirely.

Therefore, the appearance of claude-marshmallow-eap or claude-melon-eap does not establish that either model will become a publicly available Claude product.

🔬 What Developers Should Watch

#

For developers evaluating the significance of the leak, the most useful signals will come from actual API behavior rather than the codenames themselves.

Important areas to monitor include:

Model capability

#

Testing should focus on reasoning, coding, instruction following, long-context performance, structured output, and tool-use reliability.

API characteristics

#

Developers should also watch for changes in:

• Context-window limits.
• Token generation speed.
• Input and output pricing.
• Streaming behavior.
• Tool-calling support.
• Structured output support.
• Rate limits.
• Model availability.

Compatibility

#

If Marshmallow or Melon eventually becomes publicly available, developers will need to determine whether the models are drop-in replacements for existing Claude integrations or require application-level changes.

For production systems, model migration should be based on reproducible evaluations rather than anecdotal comparisons.

🌐 Implications for the LLM Market

#

The emergence of two additional Claude codenames highlights the intensity of competition among leading AI model providers.

Anthropic is competing in a market where model improvements increasingly arrive as iterative releases rather than infrequent generational upgrades. New checkpoints can target specific weaknesses in reasoning, coding, latency, agentic workflows, or conversational quality without necessarily replacing an entire product family.

If Marshmallow and Melon are indeed experimental successors to existing Claude models, their early exposure suggests that Anthropic is actively testing multiple candidates rather than waiting for a single monolithic model release.

That strategy could allow Anthropic to shorten the feedback loop between internal training, developer testing, and public deployment.

🔮 What Comes Next for Claude

#

The claude-marshmallow-eap and claude-melon-eap identifiers provide an intriguing glimpse into Anthropic’s ongoing model development, but their ultimate role remains uncertain.

The strongest current signals are:

• Two previously unannounced Claude codenames have surfaced.
• Both appear to carry the eap Early Access Program suffix.
• Community reports currently favor Marshmallow over Melon.
• Some testers reportedly consider Marshmallow’s conversational quality stronger than the current Opus 5.
• Neither model has been officially assigned to the Opus, Sonnet, or Haiku tiers.
• The shift from animal-based codenames such as Fennec to food-based names may indicate a new internal development branch.
• API exposure does not guarantee a public release.

For developers, the most important development will be whether these experimental identifiers evolve into stable, documented models with measurable improvements in reasoning, coding, agentic workflows, and API performance.

Until Anthropic officially confirms them, Marshmallow and Melon remain leaked development codenames rather than announced Claude products.