Arena says prompt caching cuts agent task costs by 260% across 26K sessions
Overview
Arena (@arena) analyzed 26,225 agent sessions from September 12 to October 23, 2026, across 30 agents and 13 labs, and reports that caching is now the main driver of agent costs.
Arena says 86% to 96% of agent tokens are re-reads of context already seen, and that without caching a task sent to one of the top-10 agents would cost 3.6 times as much, or 260% more.
Arena gives one task as an example: a typical Claude Fable 5.1 task costs $7.44 with caching and would cost $27 without it. It also reports that a single cache invalidation raised a typical task's bill by 5%. Arena says frontier labs now charge 5% of the input price for cache reads, down from full price in early 2024.
Written by AI from the articles below · updated Oct 30, 11:17 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
Arena.ai@arenaOfficialArena finds caching cuts agent task costs by 260% in session analysisAIArena analyzed over 26K Agent Arena sessions from September 12 to October 23, 2026, across 30 agents and 13 labs. It found that without caching, a task sent to one of the top-10 agents would cost 3.6x as much, or 260% more. A typical Claude Fable 5.1 task costs $7.44 with caching and would cost $27 without it. In the sample, a single cache invalidation raised a typical task's bill by 5%.

Arena.ai@arenaOfficialPickPrompt caching now drives most of AI agent costs, Arena analysis findsAIArena analyzed 26,225 agent sessions from 12 September to 23 October 2026 and found that 86% to 96% of agent tokens are re-reads of context already seen. Without caching, a typical task for a top-10 agent would cost 3.6 times as much, and frontier labs now charge 5% of the input price for cache reads, down from full price in early 2024.
Heat Trend
Not enough continuous observations to show a trend yet.