Anthropic Tests Two Mystery Claude Models

These look more like checkpoint experiments than imminent product launches, but their appearance signals that Anthropic is iterating quickly as competition around frontier models intensifies.

• –EAP indicates restricted testing rather than public availability.
• –Early testers reportedly find Marshmallow more capable and conversational than Melon, with some comparisons favoring it over Opus 5.
• –Both models reportedly spend heavily on reasoning tokens and can hit maximum-token limits, raising efficiency and cost questions for developers.
• –Strong 3D reinforcement-learning and voxel-task outputs hint at targeted capability tuning, not necessarily broad model improvements.
• –The Astra backdrop makes the timing notable, but leaked identifiers are not a reliable release roadmap.