I
Intelligence
50%
Reasoning, knowledge, coding, factuality, instruction following, and reliability.
- →Primary benchmark leaderboards and public repositories
- →Reproducible papers and published run logs
M
Multimodality
15%
Useful text, vision, image, audio, video, and document capabilities—not merely nominal input support.
- →Public capability documentation
- →Evidence of useful multimodal performance
A
Agent
15%
Tool use, planning, agentic coding, research, computer use, long-horizon completion, and recovery from errors.
- →Published agent-task results
- →Run logs and reproducible task evidence
C
Context
10%
Advertised context capacity plus demonstrated retrieval, comprehension, and long-context reasoning.
- →Public context specifications
- →Demonstrated long-context evidence
P
Price
10%
Public standard input/output pricing and cost per completed task where comparable data exist.
- →Standard public API pricing
- →Comparable task-economics evidence
Calculation
Overall score = 0.50 × Intelligence + 0.15 × Multimodality + 0.15 × Agent + 0.10 × Context + 0.10 × Price. The scorecard’s I, M, A, C and P columns are RS-normalized research-synthesis scores, not raw percentages from one benchmark.
Evidence policy & limitations
Primary evidence is original benchmark leaderboards, public repositories, reproducible papers and published run logs. Public model documentation and standard API specifications support coverage; vendor claims and aggregators are cross-checks only. Coverage may be incomplete and non-identical across frontier models, so the leading cluster should be read with roughly ±2–4 points of uncertainty.