Lentils posted example outputs produced by claude-melon-eap and claude-marshmallow-eap. The generations show strong building placement in three dimensional reinforcement learning scenes. The poster noted that both models consume large quantities of thinking tokens and repeatedly hit maximum token limits. Kimmonismus reported solid performance on medium reasoning tasks and referenced recent statements from Thariq and Boris about addressing prior criticism. The comments suggest these tests may preview updates to Opus and possibly a new Haiku model.