An Arena.ai analysis of more than 26,000 agent sessions shows re-reading context, not generating answers, drives most token usage
AI agents are mostly readers. A new analysis from Arena.ai’s Agent Arena platform found that top-performing agents process about 93 input tokens for every single token they generate.
The finding matters because it changes where the money goes. If agents spend most of their budget re-reading context, the price of that re-reading becomes the most important number on the invoice.
What Agent Arena measured
The study covered more than 26,000 sessions involving 30 AI agents from 13 labs. The data window ran from September 12 to October 23, 2026.
The headline ratio applies to top-performing agents. For every output token, they consume roughly 93 cached or re-read input tokens.
Re-reads are context the model has already seen. That means earlier instructions, prior steps, files, and conversation history that get fed back into the model as an agent works through a task.
According to the analysis, those re-reads made up 86% to 96% of total token usage. The actual output, meaning the text or code the agent produces, accounted for only about 1% to 2%.
That pattern follows from how agents operate. Unlike a single chatbot reply, an agent tends to loop: read the situation, take an action, read the updated situation, take another action. Each pass drags the accumulated context along with it.
Why caching is the whole ballgame
Prompt caching is the fix labs have leaned on. If the model has already processed a chunk of context, the provider can store it and charge less to reuse it.
Arena.ai’s numbers show how much that discount matters. Tasks run without caching cost 3.6 times more than the same tasks with caching.
One example from the analysis involved Claude Fable 5.1. A task that cost $7.44 with caching rose to $27 without it.
The pricing trend has moved in the agents’ favor. Cache discounts went from 0% in early 2024 to as low as 5% of standard input prices by 2026.
What agents are being used for
The analysis also offered a look at what people are asking agents to do. Work tasks dominated, accounting for 69% of analyzed tasks.
Work tasks are the kind that tend to involve long contexts, such as codebases, documents, and multistep processes, which are exactly the conditions where re-reading piles up.
Background: from chatbots to agents
Early conversations about AI pricing focused on output. Generated text felt like the product, so it was natural to fixate on what each response cost.
Agents flip that logic. A single task can involve many rounds of reading and acting, and the model has to keep its working memory loaded the whole time.
Agent Arena’s dataset puts numbers on that shift. With output at roughly 1% to 2% of total tokens, the old mental model of paying for answers no longer describes how agent workloads consume compute.
What this means
For businesses deploying agents, the practical lesson is to look past the sticker price per output token. The relevant figure is what a full task costs with caching applied, and whether a given workflow can actually take advantage of it. The Claude Fable 5.1 example, $7.44 versus $27, shows how wide that spread can get.
There is also a design question for developers. Agents that keep their context stable and reusable are better positioned to capture cache discounts, while agents that constantly reshuffle their inputs may pay closer to full price.
Discounts have already reached as low as 5% of standard input rates, so future gains may come less from price cuts and more from agents that simply need to reread less.
