Anthropic may already be testing its next-generation Claude models behind the scenes, and two unusual codenames are now drawing attention: “Marshmallow” and “Melon.”
According to recent leaks, Anthropic is testing models internally identified as claude-marshmallow-eap and claude-melon-eap. While the names sound harmless, the performance gap between the two could offer an important clue about where they fit in the Claude lineup.
Anthropic Could Be Testing New Opus and Sonnet Models
The most interesting detail is that Marshmallow reportedly performs better than Melon and is also said to outperform the current Opus 5.
That has led to speculation that Marshmallow could be an early test version of Claude Opus 5.1, while Melon could represent Claude Sonnet 5.1.
If that identification is correct, Anthropic may already be preparing incremental updates to both of its flagship Claude families.
The leaked naming could therefore look something like this:
- Marshmallow → Claude Opus 5.1
- Melon → Claude Sonnet 5.1
- Both models are reportedly in early-access testing
- Marshmallow is said to outperform Melon
- Marshmallow reportedly performs better than Opus 5
Of course, the codename-to-model mapping remains speculation for now. Anthropic has not publicly confirmed that Marshmallow is Opus 5.1 or that Melon is Sonnet 5.1.
Neither Model Apparently Reaches Fable-Level Performance
There is another important part of the leak that could temper expectations.
Despite the reported improvements, neither Marshmallow nor Melon is said to reach Fable-level capabilities.
That distinction matters because Fable has increasingly been associated with Anthropic’s more advanced frontier-model work. If the leaked benchmarks are accurate, these new Opus and Sonnet variants could represent meaningful upgrades without being the company’s next major capability jump.
In other words, Opus 5.1 and Sonnet 5.1 may be evolutionary rather than revolutionary releases.
What Could Opus 5.1 and Sonnet 5.1 Bring?
If Anthropic is indeed testing these models as successors to Opus 5 and Sonnet 5, users could see improvements in areas such as reasoning, coding, tool use, long-context tasks and overall reliability.
The fact that Marshmallow is reportedly already outperforming Opus 5 is particularly interesting. It suggests Anthropic may be using these internal versions to refine the next generation before a wider release.
However, early-access testing does not necessarily mean a public launch is imminent. Internal model names can change, projects can be cancelled, and experimental versions may never become consumer-facing products.
The Bigger Picture for Claude
Anthropic appears to be moving toward a broader Claude lineup where different models target different performance and efficiency levels.
An Opus 5.1 model could serve users who want the strongest reasoning and coding capabilities, while a Sonnet 5.1 model could focus more heavily on speed and cost efficiency.
That would make sense if Anthropic wants to keep improving Claude without immediately replacing its entire model lineup with a single flagship.
For now, however, Marshmallow and Melon should be treated as leaked internal model names rather than confirmed products.
If the reports are accurate, Anthropic could be quietly working on Claude Opus 5.1 and Claude Sonnet 5.1, with Marshmallow potentially representing the more powerful Opus variant.
And while neither is reportedly at Fable’s level, a model that already outperforms Opus 5 would still represent a significant step forward.
Until Anthropic officially confirms the models, though, the exact identities, capabilities and release plans remain unverified.
For continuing updates on OpenAI, ChatGPT, Codex, GPT models, and AI industry developments, bookmark:
