Amid the continuous advancement of AI conversation technology, OpenRouter has recently introduced two innovative technologies—Prompt Caching and Sticky Routing—aimed at significantly reducing the token costs of multi-turn agent conversations. The launch of this new technology means that developers can enjoy lower costs when engaging in long-term interactions with agents, thereby promoting the adoption of agent applications.

In multi-turn conversations, agents often need to repeatedly send the same system prompt, tool definitions, and strategy instructions. This leads to repeated charges even if the user changes the input only once during a six-round session. With Prompt Caching, OpenRouter can read from the cache to avoid high costs for repeated prompts. Depending on the service provider, the cost of cache reads is 0.1 to 0.5 times the normal input price, greatly reducing expenses. On the Claude Sonnet4.6 model, the cache read price is only $0.30 per million tokens, while the normal input price is $3.00.

image.png

To ensure this mechanism remains effective in multi-turn conversations, OpenRouter's Sticky Routing technology directs subsequent requests to the same provider that holds the hot cache, allowing developers to continue benefiting from low-cost cache reads across different conversation rounds.

However, to fully utilize this innovative technology, developers must ensure they maintain the same session throughout the process. For cache effectiveness, OpenRouter recommends users set a session_id, which can immediately enable Sticky Routing after the first successful request, not just after a cache hit. Additionally, developers can easily check whether the cache is working by reviewing the usage data in the request response.

As AI technology continues to advance, this new technology from OpenRouter not only provides developers with a powerful tool for cost control but also opens up new possibilities for agent application scenarios.