Survey Report 2026 · March

LLM Usage
& Ratings
Survey Report

48 real users from mainland China, the U.S., and Europe shared their LLM usage habits, scenario preferences, and subjective ratings. Data collected March 2026.

0
Valid responses
0
Models mentioned
0%
Daily heavy users
0%
Report some dependence
01 · Demographics
Respondent Profile
From academia and the workplace, spanning China and the U.S. — a typical cohort of heavy LLM users
Gender Distribution
♀ Female 50% ♂ Male 48% — Undisclosed
Regional Distribution
🇨🇳 Mainland China 58% (28 respondents) ·  🇺🇸 United States 31% (15 respondents)
Age Distribution
Age 32–3538%
Age 26–2823%
Age 36–4015%
Age 29–3110%
Age 22–258%
Age 41+6%
Occupation
Traditional Industry29%
PhD Student27%
Tech Industry13%
Research Asst. / Postdoc6%
Freelancer6%
Professor4%
02 · Behavior
Usage Behavior Analysis
High-frequency, dependent, and willing to pay — this cohort has fully integrated LLMs into their daily workflow
Usage Frequency
Dependency Level
Cross-Verification Habits
🔥 Hidden Finding · Usage Frequency × Dependency
The More You Use, the More You Depend——those using countless times daily show 52% strong dependence
Among those using a few times a week, only 11% report strong dependence. Higher frequency = higher switching cost.
52%
Countless/day
→ Strong dependence
31%
Few times/day
→ Strong dependence
11%
Few times/week
→ Strong dependence
Max Monthly Willingness to Pay
💡 19% refuse to pay at all; 68% cap at under $20/month
Gender × Payment Willingness
💡 Males show higher willingness to pay, some up to $200–300/month; females cluster at $0–20
03 · Model Landscape
Model Landscape
Striking differences between mainland China and U.S. users — toggle to explore
View by region:
Models Used (All, n=48)
Payment Status
Model Popularity Bubble Chart (bubble size = no. of users)
💳 Paid vs. Free User Behavior
Paid Users (n=33)
Countless/day61%
Few times/day27%
Few times/week12%
Free Users (n=15)
Countless/day33%
Few times/day27%
Few times/week33%
🧤 Paid users's "countless/day" rate is nearly that of free users — those who pay, use more intensively
04 · Scenarios
Scenario Preferences
ChatGPT leads across all scenarios; Gemini tops coding, and Claude earns outstanding reviews in dev circles
Scenario 🥇 Top Pick 🥈 2nd Pick 🥉 3rd Pick
💬 Daily Search / Casual Chat ChatGPT 19 votes Doubao 16 votes DeepSeek & Gemini 14 votes
☕ Deep Research / Academic Writing ChatGPT 21 votes Gemini 20 votes DeepSeek 11 votes
💻 Coding / Debugging Gemini 22 votes ChatGPT 14 votes Claude 13 votes
📄 Long-form Text / Reading Papers ChatGPT 23 votes Gemini 17 votes Doubao & DeepSeek 8 votes
✍ Creative Writing / Brainstorming ChatGPT 24 votes Gemini 19 votes Doubao & DeepSeek 9 votes
Model Vote Distribution by Scenario
🎓 Interesting: Coding Scenario Preferences by Occupation
PhD Student (n=13)
Gemini54%
Claude38%
ChatGPT31%
Tech Industry (n=7)
Gemini71%
ChatGPT43%
Claude14%
Traditional Industry (n=15)
Non-coders47%
Gemini40%
Claude20%
05 · Ratings
Model Capability Ratings
Rating scale 1–5; 'don't know' and 'hard to evaluate' excluded. Sample size (n) shown top-right of each card.
ChatGPT n = 27–36
Claude n = 9–16
Gemini n = 23–31
DeepSeek n = 15–28
Qwen n = 4–10 ⚠ Small sample
🤥 Hallucination Rate (1=rarely, 5=often; lower is better)
Claude
2.50
Best
Qwen
2.60
Good
Gemini
2.96
Average
ChatGPT
2.97
Average
DeepSeek
3.10
High
🖼️ Image Generation (Gemini Far Ahead)
Gemini
4.20
Far Ahead
DeepSeek
2.80
Average
Qwen
2.75
ChatGPT
2.54
Claude
2.44
Gemini image generation avg. 4.20 — 1.4 points ahead of 2nd place
06 · Dev Tools
AI IDE & Plugin Usage
Cursor, GitHub Copilot, and Claude Code form a three-way race; Claude Code repeatedly praised as standout
🖱️
Cursor
12
25% usage rate
🐙
GitHub Copilot
11
23% usage rate
Claude Code
10
21% usage rate · ⭐ repeatedly praised as standout
🤖
Non-coders
6
12.5% of respondents
07 · Hidden Insights
6 Hidden Findings
Click each card to expand details
08 · Voices
What Users Think About LLMs
I just want to lie down and have it do all the work for me!
Seeing how AI works through problems step by step — that really impressed me.
Claude Code — truly impressive.
Gemini can read browser content directly — that's really cool.
There are too many apps now; none feels truly mind-blowing yet.
Agents are about to replace me.
I hope agents slow down their development.
Accuracy still needs improvement.
Often logically confused — confidently making things up.
They all feel about the same to me.
Impressive.
Really great — very practical.
Some are still unstable. I paid, but there are usage limits.
Sometimes it hallucinates. But mostly impressive.
Technology changes life.
If only it could generate files with one click.
Hallucinations are still quite noticeable.
Nothing much — just want to learn more about LLMs.
When can I just lie in bed and have AI read my mind and get things done?
Current LLMs still lack true creativity — they can't solve problems humans can't.
Better data filtering and source verification needed — some official sites can't be read.
LLMs demand precise prompts — the more detailed the question, the better the answer. It'd be great if they could help users ask better questions.
Since my use cases are fairly simple, I don't notice much difference between models. I'm not sure what each excels at — I only use the most basic features.
Even if LLMs bring some convenience, I believe the drawbacks will outweigh the benefits in the long run — like many modern technologies. There's a tendency to deprive people of independent thinking and foster intellectual laziness. Use carefully. Convenience comes with consequences.
Seems convincing, but when you get into your own field, you realize it's just stitching together nonsense.
Wish Gemini had a Projects feature. When will Codex get a Windows version or support older MacBooks?
Hard to judge. Solves many quick problems, but also creates more. Output quality needs more comprehensive evaluation.
A bit dumb.
Faster multimodal end-to-end consumer deployment; full-duplex model-side interaction rollout, please.
We need an open-source operating agent.
It's great — incredibly helpful.
A new era has begun.
Sometimes the information sources feel unreliable.
Future potential seems limitless — another leap in technology.
All pretty good — should keep getting better.
Too expensive