Token Consumption as a New Way to Judge an AI Application

Originally published in Chinese on August 6, 2024. Applications built on models now carry a new variable, the token. This piece asks whether token consumption can serve as one more yardstick for application value, next to the metrics that internet and mobile internet products have long used. The traditional metrics In the internet and mobile internet eras there were clear reference metrics. They include user scale and quality (active users, new users, composition, retention, total active days), engagement (launches, time spent, pages viewed, interval between uses), channels (new and active users on Android and iOS, plus anti-fraud metrics), features (feature activity, page paths, conversion), user attributes and profile analysis, and revenue analysis. These are mature metrics from the past, and they still apply today. Tokens as a new dimension Now that AI applications carry model and token parameters, evaluating application value may open a new dimension, with tokens as the reference unit. To a degree, the number of tokens consumed becomes a reference metric for value, though more consumption is not always more value. T