Top ten agents in Agent Arena achieve an input-to-output ratio of 93:1

Based on our analysis of over 26K sessions, top ten agents in Agent Arena process roughly 93 input tokens for every output token generated. This is why labs across the industry are racing to reduce cached input costs. The video below demonstrates how caching works in practice using real sessions from @AnthropicAI's Opus 5.5 and @GoogleDeepMind's Gemini 4 Argon.