Claude API Now Supports Mid-Conversation Tool Changes Without Invalidating Prompt Cache
Anthropic's Claude Platform has introduced the ability to add or remove tools mid-conversation without invalidating the prompt cache. Additionally, the platform now features fallback routing, which automatically redirects classifier-blocked requests to a recommended model. This update significantly reduces latency and API costs for developers building complex agentic workflows that dynamically swap tools. It prevents the need to reprocess long system prompts or conversation histories when tools are modified. By preserving the prompt cache during tool changes, developers can maintain fast response times even as the agent's capabilities shift. The new fallback routing mechanism ensures system resilience by automatically handling requests blocked by safety or content classifiers.
## BACKGROUND
Prompt caching is a technique where the LLM provider saves the processed state of repeated prompt segments, such as system instructions or long contexts, to avoid redundant computation. Tool use, or function calling, enables LLMs to interact with external databases and APIs by generating structured arguments. Previously, modifying the set of available tools mid-session would alter the prompt structure, forcing the system to invalidate the cache and reprocess the entire input.