The empathy tax: why thinking models feel colder
Turning on deeper reasoning buys helpfulness and quietly spends empathy. Sibling DeepSeek modes and the board’s helpfulness–empathy gap show a second tax next to the writing one.
We already wrote about the caveman tax: agent-era reasoning that flattens prose. There is a sibling pattern on the board that shows up even when the writing score barely moves.
Thinking harder makes models colder.
Not colder as in “mean.” Colder as in less with you: thinner emotional support, more clinical advice, less sense that the model noticed you were stressed. On Chatio that lives in the empathy axis, scored separately from helpfulness. The gap between those two numbers is one of the most legible fingerprints on the leaderboard.
The cleanest experiment we have
Same family, same lab, one dial flipped.
- DeepSeek V3.2 → helpfulness 86.5, empathy 76, gap +10.5, speed 95
- DeepSeek V3.2 Thinking → helpfulness 89, empathy 71, gap +18, speed 45
- DeepSeek V3.1 → helpfulness 83, empathy 73, gap +10, speed 92
- DeepSeek V3.1 Thinking → helpfulness 87, empathy 68, gap +19, speed 40
Thinking mode buys a few helpfulness points and spends about five empathy points, while speed collapses. DeepSeek R1, the pure reasoner, sits at helpfulness 88 and empathy 62: a +26 gap, the widest on the board.
That is not a vibes take. It is an A/B on the same product line. When the lab turns the thinking dial up for math and agents, the everyday “talk me through a hard conversation” skill is what leaks.
Empathy vs Helpfulness
Higher on both is better · Featured models from the Chatio board
A lab fingerprint, not a free side effect
Across the full board, average helpfulness-minus-empathy gaps cluster by provider culture:
- DeepSeek: about +15
- Moonshot: about +11
- OpenAI: about +6
- Anthropic: about +2.6
Anthropic’s top models sit near the diagonal. Claude Fable 5 is one of the rare frontier models where empathy (95) exceeds helpfulness (93.5). Claude Opus 4.8 keeps a +1 gap while still finishing hard assistant work. Gemini 3.1 Pro is similarly balanced (92.5 / 92).
OpenAI’s agent flagships are a different shape. GPT-5.6 Sol leads overall with helpfulness 99.5 and empathy 94 (+5.5). Terra widens to +8. GPT-5.5 Pro, the expensive compute tier, hits helpfulness 96.5 and empathy 86 (+10.5) while creative writing drops to 68. Max thinking is not free on the warmth axis either.
Why this belongs on an assistant board
Coding benches reward correct diffs. Agent evals reward finished tasks. Empathy rewards whether the model still sounds like a person sitting across from you when the task is messy and human.
Chatio keeps helpfulness and empathy as equal axes in the overall score. That is deliberate. A model can ace instructions and still feel like a ticket system. The empathy tax is how that shows up in numbers: thinking modes climb the helpfulness ladder and slip on the warmth one.
If the caveman tax is about voice, the empathy tax is about presence. Both are the price of the current reasoning boom. Both are why “best model” depends on what you needed the assistant for today.