Capability scorecard · Five public measures
Reasoning
Performance on difficult, research-level mathematical problems.
GPT-5.5 Pro pre-release (high) OpenAI
Kimi k2.6 Moonshot AI
Higher scores show stronger hard-problem solving; they do not establish broad reliability.
Software work
Resolution of verified, real-world software engineering issues.
Claude Opus 4.7 (max) Anthropic
GLM 5.2 (max) Zhipu AI / Z.ai
Measures bounded repository tasks, not unattended ownership of production systems.
Multimodal
Understanding across short, medium, and long video without subtitle assistance.
Gemini 1.5 Pro Google DeepMind
video-SALMONN 2+ ByteDance
Captures one form of visual-language understanding, not embodied world modelling.
Autonomy
The length of software tasks models complete at an estimated 80% success rate.
Claude mythos Preview early Anthropic
Public dataset has no positive scored record.
Longer horizons matter, but a benchmark task is not the same as safe, persistent agency.
Generalization
Adaptation to novel abstract visual reasoning tasks.
GPT-5.6 Sol (Max) OpenAI
DeepSeek V4 Flash 0731 (Max) DeepSeek
Strong performance can reveal flexible reasoning, while benchmark saturation can weaken the signal.
Optional plain-language interpretation
Optional interpretation
The model receives the selected public record, not an open web prompt. It cannot alter the score.