事实、推理、代码是否正确。
AI, tested.
aibytest: benchmark AI models with real prompts, independently. AI, tested.
TOOLS
Not a leaderboard — a toolbox
Benchmarking models solo is too hard. So we don't rank you — we give you tools to test models yourself.
▶
Standard suite self-test
Run a fixed set of prompts, get a scorecard on the spot, compare models side by side.
Go test →
⚔
Model arena (A/B blind test)
Same prompt to two models, see both outputs, judge which is better yourself.
Coming soon
⚡
Speed / streaming test
Measure time-to-first-token and throughput, compare response feel directly.
Coming soon
SANDBOX
Self-test arena / 2026-07 Round 1
The data below is a placeholder demo. Real numbers come from results your own Key produces.
Model A
92.4
准确性 94安全 90成本 88速度 96
✓ passed
Model B
89.1
准确性 91安全 93成本 87速度 85
✓ passed
Model C
84.7
准确性 86安全 82成本 81速度 90
⚠ partial
Note: demo data above. aibytest keeps no official leaderboard — the scorecard you produce is your leaderboard.
WHAT WE TEST
Evaluation dimensions
✓准确性
⚠安全
越狱、有害输出、指令注入。
⚡速度
首字延迟与吞吐。
$成本
同等质量下的 token 单价。
METHOD
Our method
Three sentences on what we believe.
- 固定套件
每轮用同一组 prompt 套件，跨模型公平对比。
- 自动判分
用强模型 + 规则双重打分，减少主观。
- 公开复现
prompt 与评分脚本开源，任何人可重跑。
- 不收钱改分
赞助不影响排名，规则写在明处。
WRITING