Unreleased Claude AI Models Spotted in Anthropic Developer Pipeline

Tech enthusiasts on social media have uncovered references to unreleased artificial intelligence models hidden within Anthropic's developer infrastructure. According to leaks shared on X by developer Lentils, two experimental systems codenamed "claude-melon-eap" and "claude-marshmallow-eap" are actively being trialed in an early access testing phase.

Initial test outputs suggest these upcoming releases place a heavy emphasis on extended internal reasoning before answering, frequently using up high token limits to process complex prompts. Early comparisons shared online claim the "marshmallow" variant delivers noticeably smoother conversational interactions than Anthropic's existing flagship models, alongside improved spatial capabilities like arranging objects in 3D environments.

It remains uncertain when or if these specific builds will be made available to the general public, as tech companies regularly test internal prototypes that never reach official launch. However, this backend activity indicates that Anthropic is actively pushing toward deeper reasoning capabilities to stay competitive with rivals who are heavily investing in reasoning-first AI models.