01
Reasoning
FrontierMath
+13.4pts
Western lead
Performance on difficult, research-level mathematical problems.
Western
52.4%
GPT-5.5 Pro (high) OpenAI
Chinese
39.0%
Kimi K2.6 Moonshot AI
Scale: 0–100%
Capability scorecard · five public measures
No side leads every kind of intelligence.
- 01Reasoning West +13.4
- 02Software work West +4.8
- 03Multimodal China +4.7
- 04Autonomy No comparison
- 05Generalization West +22.1
3 Western leads1 Chinese lead1 No comparison
Best recorded result, not average ecosystem quality. Each panel compares the highest positive dated score in that benchmark. Different panels use different units and must not be added together.
02
Software work
SWE-bench Verified
+4.8pts
Western lead
Resolution of verified, real-world software engineering issues.
Western
83.5%
Claude Opus 4.7 (max) Anthropic
Chinese
78.7%
GLM-5.2 Zhipu AI / Z.ai
Scale: 0–100%
03
Multimodal
Video-MME
+4.7pts
Chinese lead
Understanding across short, medium, and long video without subtitle assistance.
Western
75.0%
Gemini 1.5 Pro Google DeepMind
Chinese
79.7%
video-SALMONN 2+ ByteDance
Scale: 0–100%
04
Autonomy
METR task horizon (80%)
—
No direct comparison
The length of software tasks models complete at an estimated 80% success rate.
Western
3.1hr
Claude Mythos Preview (Early) Anthropic
Chinese
No comparable result
—
Public dataset has no positive scored record.
Scale: 0–4 hours
05
Generalization
ARC-AGI-2
+22.1pts
Western lead
Adaptation to novel abstract visual reasoning tasks.
Western
95.0%
GPT-6 Astra (max) OpenAI
Chinese
72.9%
DeepSeek V4.1 Flash (max) DeepSeek
Scale: 0–100%
Evidence desk
Ask what a score means before deciding what it proves.
Answers are written from this site’s method notes and the current snapshot. The desk explains the record; it can’t change the score.
Pick a question, or type your own about a benchmark, a score, an access label or a method term.