93 to 1: For Every Token an Agent Writes, It First Reads 93

Arena.ai analyzed 26K+ sessions, and the input-to-output ratio of the Top 10 agents is roughly 93:1. For every 1 token generated, 93 input tokens have to be processed first. That ratio explains why labs are racing to cut the cost of cached input.

The way agents work makes it hard for them to read anything just once. Multi-turn conversations, tool calls, error retries — every step shoves the historical context and tool results back into the model. One commenter noted that every tool call drags the bash history along with it. It sounds like a joke, but it lines up with 93:1. Agents spend most of their time re-reading homework they just wrote themselves.

Arena.ai demonstrated how caching works using real sessions from Opus 5.5 and Gemini 4 Argon. There is a detailed article below the video by @DawidGalarowicz. The original thread is here, and the article is here.

Caching is not a new concept, but it is turning into a separate pricing tier. One commenter brought up the 5-minute default expiration: step away to eat a sandwich and your bill could double. That claim comes from a reply, not a general conclusion, but it does highlight how cache retention time affects cost.

More concrete numbers come from commenter sree.world. Grok 4.7 re-reads about 89% of its tokens, and that accounts for 65% of the bill; Fable 5.1 re-reads about 92%, yet that accounts for only 12% of the bill. According to that reply, xAI charges 25% of the input price for cache reads, while the others charge about 5%. The re-read ratios are close, but the share of the bill differs by several times — the price rate is what is doing the work. The data and attribution here come from that commenter, not from conclusions in Arena.ai's original article.

The chart puts the cache re-read ratios and cost shares of different models side by side. When you look at this chart, don't just look at the blue bars. The gray bars are the bill. It is not surprising for an agent to re-read about 90% of its tokens; what is surprising is that, with the same re-reading, some models can push the cost down to 12%, while others have to pay 65%.

Another commenter asked whether cache retention limits under long histories would change which agent is the cheapest. That question is more practical than looking at the input unit price alone. The longer the session, the more often caches expire and get re-read, and those small numbers on the price sheet get amplified.

So when choosing an agent, the cache hit rate, expiration time, and cache-read unit price may do more to determine your end-of-month bill than the output unit price. 93:1 is only the surface ratio; behind it is a quiet war over cache pricing.