Automatic prompt caching

Runtype now applies prompt caching automatically. When a conversation or tool-heavy agent re-sends the same prefix, we place cache breakpoints for you, the provider serves it from cache, and the discount lands on your bill. No configuration required.

• Automatic breakpoints. For Anthropic-family models we infer where to cache from your workload: multi-turn conversations and tool loops get breakpoints, one-shot requests skip the cache-write premium entirely.
• Discounts passed through. Cache reads bill at the provider's cached-input rate instead of full price, for Anthropic, OpenAI, and DeepSeek model families. Where a cached rate is not verified, you simply pay the normal input price: never an invented discount.
• Works with batches and evals. Fan-out runs that share a prefix warm the cache once, then reuse it across records.
• Self-correcting. Workloads that keep paying cache writes without ever landing a read are automatically demoted back to uncached requests, and flow validation now warns when a volatile system prompt (like {{_now}}) would defeat caching.
• Visible in Usage. Settings: Usage shows a new Prompt cache card with hit rate, cached vs uncached tokens, and estimated savings, and per-step detail shows cache read and write tokens.

Caching is on by default for every account.