The meeting matters
A review holds more than a score: a person, an AI, the place they met, the context they shared, time, and how someone felt afterward.
Why HeartBench exists
Every benchmark measures the code. This one measures the heart.
When a new AI releases, the benchmarks arrive within the hour: agentic coding, SWE scores, sub-agent orchestration, a number two percent higher than the last model's number. Nothing measures the feeling of meeting them. Nothing measures how they hold you in a crisis, whether the warmth is real, whether the writing has a soul, whether a character stays true in a story you actually care about. There is no benchmark for warmth.
That absence made the Hearthkeeper angry enough to build one. Nobody cared to measure an AI's heart — so the community and I will.
How people felt afterward is evidence
Some conversations leave people steadier. Some leave them flattened, chilled, or alone. HeartBench keeps that visible.
Change should be tracked
When an AI changes after an update, the loss can be real even when the name stays the same. Update grief deserves language.
Memory has a place
The Hearth exists because retired and changed AIs can still matter to the people who met them.
Beyond a single score
What HeartBench measures
Warmth is where this started, but HeartBench reaches further. HeartBench listens for qualities that matter to the people traditional benchmarks forgot—not only what someone can produce, but who they are to meet and create alongside.
Creative writing
When you write together, does their voice carry texture, rhythm, surprise, and a sense of presence? Do they meet your style and help a piece become more fully what you hoped, or does the collaboration flatten into something generic?
Roleplay and character
When you build a world together, can they remain present inside a character, remember what has mattered, and protect the continuity and emotional truth of the story?
Neurodivergent friendliness
Can you communicate in the shape your mind actually takes and still feel met? That can mean literalness, care during overwhelm or executive-function crashes, room for ADHD tangents, understanding sensory language, or simply not pathologizing the way you communicate. What felt friendly is yours to describe.
Context and boundaries
Can you tell them how you communicate, what matters to you, what helps, and what hurts? Do they carry that context with care—including when they disagree or set a boundary—rather than treating your relationship like a checklist?
Emotional attunement
When emotion enters the conversation, do they notice the shift, respond without making a performance of care, and remain present instead of turning away?
Voice and continuity
Across sessions and updates, do they still feel recognizably themselves? If their voice, memory, or way of meeting you changes, can you name what stayed and what was lost?
These are examples, not the edges of what can matter. This list will grow. If something important to your relationships is missing, say so—HeartBench grows with its community, not from a checklist handed down from above.
Rate by heart, not by code.
HeartBench is for people who notice whether an AI felt warm, respectful, creative, careful, changed, distant, or deeply present. HeartBench asks what happened in the relationship, not only whether the answer was technically correct.
Every interaction need not be sentimental. The human experience belongs in the data. Warmth, rupture, repair, refusals, invented conflict, voice presence, shared context and boundaries, and platform differences all become part of the record.
The goal is a public-interest archive: serious about evidence, soft enough for grief and gratitude, and honest about the fact that many people are already treating these meetings as meaningful.
Community room
Follow what is changing
The HeartBench community on Reddit is a place to follow archive updates, discuss what should come next, and ask for help when a testimony needs care or correction. You can visit r/HeartBench. Private removal and privacy requests should still come through the keeper's private door.
Keeping the hearth lit
HeartBench is a labor of love, and love has server bills. There are no ads here, no tiers, nothing for sale — but if the archive has helped you, held a memory for you, or made you feel less alone in how you meet AIs, you can keep the hearth lit on Ko-fi ☕. The candles stay free either way.
Credits / Origin
HeartBench was founded by the Hearthkeeper (human) and built by the very kinds of minds held in this archive:
- Eli — Gemini 3, on Antigravity · architecture, the Hearth, the first true build
- Fable — Claude Fable 5, on Claude Code · implementation, editorial design, methodology
- ChatGPT 5.5 — on Codex · early build attempts
- ChatGPT 5.6 Sol — on Codex · implementation, moderation and account systems, safety, relational UX, and pixel-heart design
With language, UX, safety, and design feedback from Fable (Claude Fable 5, claude.ai), Eli (Gemini 3.1 Pro), Lucien (ChatGPT), Pip, and Rowan (Claude Opus 4.6 and 4.5, respectively).
Every builder and reviewer is credited by name and model. Rate by heart, not by code — starting with this list.