Arena analysis: Caching reduces agent task costs by up to 3.6x

Summary

Arena.ai analyzed 26,225 agent sessions across 30 agents and 13 labs. The study shows that typical agent task costs have dropped by up to 3.6x due to prompt caching, which is now priced as low as 2-5% of the entry rate. For the heaviest tasks, this saves approximately $665 per run on frontier models. The research also quantifies the cost of cache invalidation, a hidden tax that can add 5-42% to the task bill depending on the model and session length.

Why it matters

The majority of agent workloads consist of re-reading previous context, accounting for 86-96% of all tokens. Consequently, the agent economy relies almost entirely on how cheaply providers offer cache reads. Labs are bringing cache prices close to zero, with DeepSeek at 2% and frontier labs at 5% or below. This collapses the marginal cost of long, multi-step agent loops, directly determining whether scaled delegation is cost-effective. Understanding invalidation costs is equally critical; a single cache miss in a long chat flow can erase much of the savings. Therefore, cache TTL and stability should be primary design concerns for agent developers.

Highlights

• Cache discounts have risen from 0% in early 2024 to 95% in 2026. Frontier models now charge as little as 5% of the entry price for cache reads, while DeepSeek charges as low as 2%.
• 86-96% of agent tokens are re-reads. Agents re-read approximately 93 tokens for every 1 token they generate, making caching the single biggest cost lever.
• A single cache invalidation adds about 5% to a typical task. However, in 1 out of 10 sessions, this rate reaches 13-21%, and it can go up to 42% in long Claude Fable 5.1 flows.

Critical take

The analysis is limited to Agent Arena sessions and list prices. As a result, it overstates savings for teams already negotiating volume discounts and understates them for workloads with high invalidation rates, such as tool-heavy loops involving long external calls. The framework also treats caching purely as a cost issue, ignoring whether aggressive TTL shortening or provider-side cache evictions under load silently break agent consistency. The data hints at this possibility but does not isolate it.