Top 10 Arena Agents Reread About 93 Cached Tokens per Output Token

Across the top 10 agents on Arena.ai’s Oct. 8 leaderboard, cached context was reread about 93 times for every token produced, a ratio that makes cache pricing a major part of the cost of multi-step AI tasks.

The sessions ended between Sept. 12 and Oct. 23, 2026, and involved 30 agents across 13 labs. The analysis covered 26,225 sessions, with a cap of 1,000 sampled sessions per model. The top 10 were defined by Arena’s Oct. 8 leaderboard.

Cached re-reads made up 86% to 96% of token use, while model output accounted for 1% to 2%. A re-read is context the model received during an earlier call. The ratio describes reuse of previously seen context in these sessions, not all input tokens used by AI agents generally. The output measure includes reasoning tokens.

“Across the top 10, agents re-read about 93 tokens for every one they write,” wrote Dawid Galarowicz, author of the analysis.

Multi-step agents may make several calls while working through one request. Later calls can carry forward the conversation and results from tools used earlier, even when that information has already been provided to the model. A provider can serve recurring context from a cache at a lower rate than new input. If the cache expires or the context changes, the material may need to be sent again at the full input rate.

That difference changes modeled task costs even when the number of tokens stays the same. A cost model using list prices and holding token use constant put a typical top-10 task at 3.6 times the cost when cached tokens were billed at the full input rate.

For Claude Fable 5.1, the model placed a typical task at $7.44 with caching and $27 without it. The figures are modeled costs, not customer bills. A single cache invalidation, which requires previously cached context to be sent again, added about 5% to the typical task cost in the model.

The sessions came from Agent Arena, a platform where people use agents for work such as coding, research and document creation. The results describe activity on that service and do not establish that agents elsewhere have the same token mix or costs.