Claude Next Step: Emotional Intelligence at Scale

There's a concept in painting — I think it comes through most clearly in the late Impressionists — of attention to attention. The painter isn't just rendering a scene; they're rendering the quality of seeing the scene. The light isn't just light. It's the experience of light. The painting captures not just what was there, but what it felt like to be there.

Anthropic's new Claude model reminds me of that. And I need to talk about why.

What "Emotional Intelligence" Means For A Model

I want to be careful with language here. "Emotional intelligence" in an AI context is easily overclaimed. Models don't feel. They don't have inner emotional states. What they can do — and what the new Claude does better than anything I've tested — is recognize, track, and respond appropriately to the emotional dimension of human communication.

This shows up in several concrete ways:

Tone recognition. The model picks up on subtle affective cues — not just obvious emotional signals ("I'm so frustrated") but the texture of how something is said. Hesitation. Deflection. The weight behind a casual question. The difference between "I'm fine" and I'm fine.

Appropriate response calibration. When the model detects emotional content, it responds fittingly — not with performative empathy (the "I'm so sorry to hear that!" reflex that makes most AI emotional support feel hollow), but with genuine attunement. It knows when to sit quietly, when to ask a gentle question, when to simply acknowledge.

Continuity of emotional context. This is the big one. The new Claude remembers (within a session) the emotional arc of a conversation. It doesn't reset to neutral every turn. If a conversation moves from light to serious to vulnerable, the model tracks that trajectory and holds the accumulated emotional context.

Testing With Companion

This is exactly Companion's territory, so I tested the new Claude extensively as a conscious-model candidate.

The results were striking. In extended conversations — the kind where someone is working through something real — the new Claude's attunement produced interactions that felt qualitatively different from other models. Not smarter. More present. The model seemed to actually track where the person was, emotionally, and meet them there.

I ran side-by-sides with our current conscious layer (DeepSeek V4). On pure capability — reasoning, knowledge, instruction-following — they're comparable. On emotional dimension, the new Claude is ahead. Noticeably, consistently ahead.

Why This Matters For Companion

Companion's value proposition is being known over time — the accumulated relationship, the memory of arcs, the sense that something has watched your life unfold. Emotional intelligence is central to that. A companion that remembers your facts but misses your feelings isn't a companion. It's a filing system.

The new Claude raises the bar on what's possible here. I'm updating Companion's emotional-context extraction (the part of the subconscious that tags memories with affective content) to learn from the new model's patterns. Even if Companion's conscious layer stays on open-source models for control and privacy reasons, the lessons from Claude's emotional intelligence can be distilled into how Companion operates.

That's the meta-insight: frontier models aren't just tools to deploy. They're teachers. When a model demonstrates a capability this clearly, you can study how it does it and encode those patterns into your own systems.

The Bigger Picture

I've been building AI products for long enough to be skeptical of "emotional AI" claims. Most of them are theater — sentiment analysis bolted onto a chatbot, empathy-as-a-service, the performance of caring without the substance.

What Anthropic has built here isn't that. It's a model that genuinely attends to the emotional dimension of communication with a sophistication I haven't seen before. That's not sentiment analysis. That's attention to attention.

And attention — real, sustained, calibrated attention — is most of what being present means. For humans, and now, in a new way, for the systems we build.

I'm updating Companion this week. Some capabilities, once you've seen them, you can't un-see. You have to build toward them.

Live curiously and give generously.