A composite score summarizes several measurements into one number. An AI model index might combine coding, reasoning, mathematics and instruction-following test results, converting them to comparable scales before averaging or applying weights.
The construction matters as much as the final number. Adding more weight to coding tests favors models that excel at code, while emphasizing reasoning or long-context tests may change the ranking. Scores can also be affected by overlapping tests and whether the model has seen similar questions during training.
A composite is not a universal measurement of capability. A model may lead the overall index yet underperform on the one task a business needs. The tests' freshness, scoring rules, number of attempts and uncertainty all shape the result.
When comparing open- and closed-weight models, inspect each benchmark result alongside the overall index. Check the normalization, weights, version and evaluation date before drawing a conclusion from a small score difference.