Elephant Alpha lands mixed EQBench v3 scores

My read is that Elephant Alpha looks like a sharp, structured model rather than a naturally warm one: it can reason through emotionally loaded prompts, but it does not turn that into standout human rapport.

• –The 8.2 analytic score is the clearest signal here; the model seems better at parsing scenarios than performing empathy theater.
• –The 4.3 human score, paired with a 2.6 sycophancy note, suggests it resists glazing but may feel flatter than chat-first models.
• –A 5.4 moral score is respectable for a stealth model and keeps it from looking one-dimensional.
• –The gap to DeepSeek-V3-0324 is small enough that this reads as incremental progress, not a leaderboard breakout.
• –EQ-Bench v3 is Opus-judged and Elo-based, so the absolute numbers matter less than the shape of the profile.