Prompt caching

Providers charge much less for prompt text they have seen recently. Sovereign keeps that discount working.

What Sovereign does

• Keeps conversations on one provider, so the cache built on earlier turns is used on the next one.
• Routes with the cache in mind: what a provider already holds of your conversation counts when Sovereign decides where the next turn goes.
• Turns caching on for Claude models when your request doesn't mark any cache_control. Your own markers are kept as they are.
• OpenAI models cache on their own.

What you can do

The cache only covers the part of a prompt that is identical to before, from the first character. Anything that changes near the start makes the whole prompt full price again.

• Keep the start of the prompt the same between calls. Put dates, times, request ids and other changing values at the end of the latest message, not at the top of the system prompt.
• Send tools in the same order every time.
• Don't rewrite earlier turns unless you mean to: compaction and edits start the cache over.

Your dashboard's Insights page shows when a prompt keeps changing at the start, what it costs, and where it happens.