Agent Arena finds AI agents read about 93 tokens for every one they write

Arena.ai's Agent Arena found top AI agents process about 93 input tokens per output token, making prompt caching the key driver of costs.

Published: 2026-11-02