OpenRouter threads now cache their prompts

Every round of an agent thread re-sends the same large prefix to the model: the agent's instructions, its tool documentation, and the whole conversation so far. Some providers cache that prefix on their own, but several of the models you can run through OpenRouter — Anthropic's Claude among them — only cache when the request explicitly asks for it. Ours never did, so those threads paid full input price for the same prefix on every single round.

Now every chat request sent through an OpenRouter provider opts into prompt caching. After the first round, the model re-reads the unchanged prefix at the provider's reduced cache rate — roughly a tenth of the normal input price on Claude — which adds up quickly on long, tool-heavy threads. Providers that cache automatically are unaffected, and short prompts below the provider's minimum cacheable size simply keep behaving as before.

You can see it working: the thread's token chip and usage breakdown now show how much of the input was served from cache.