Anthropic New Marshmallow Model Beats Opus 5, First Major Update Since Fable 5

Anthropic has introduced two new internal language models to its early access program to test potential updates to its Claude AI family. Testing of claude-marshmallow-eap and claude-melon-eap indicates that Marshmallow outperforms Melon and provides a superior conversational experience to Opus 5, though neither model matches the capabilities of Fable.

The EAP suffix identifies internal checkpoints used for evaluation before a public deployment. This naming convention was previously used for the Fruitcake and Honeycomb models prior to the launch of Fable 5, suggesting a forthcoming release for the Opus or Haiku lines.