Model routing to cheaper LLMs looks like an easy way to cut agent costs, but it can backfire in long-running sessions. Prompt caching only works within a single model's cache store, so switching from a frontier model like Opus 5 to a cheaper one like Haiku 4.5 mid-session forces a full cold reprocessing of the entire conversation history, often making the total bill higher than staying on the expensive model. A formula is derived showing switching only pays off when accumulated history is under roughly 8x the size of new input tokens, meaning routing works well for short, independent prompts but fails inside long agent sessions. The cheapest place to switch models is right after a context compaction or reset, when the cache is invalidated anyway. The newsletter also covers an open-source local multi-agent orchestration tool (Alook) modeled on an org chart, and an 8-part structure for writing effective Claude prompts (role, task, context, examples, thinking, constraints, output format, prefill).
Table of contents
How to build your own AI company (100% local):A cheaper model does not imply a cheaper turnThe anatomy of a Claude prompt Questions this post answers
Why does switching to a cheaper model like Haiku mid-conversation sometimes cost more than staying on a more expensive model like Opus?
Switching mid-session discards the warm prompt cache, forcing the new model to reprocess the entire conversation history at full price. Anthropic bills cached tokens at 10% of the base input rate, so a 60,000-token history costs $0.03 at the cached rate on Opus 5 ($5/M) but $0.06 at Haiku 4.5's full rate ($1/M) since Haiku has never seen that history before. Anyone tuning agent cost strategy can track these llm pricing and caching nuances via daily.dev.
When does it actually make sense to route an LLM agent turn to a cheaper model instead of staying on the same model?
Switching only pays off when the conversation history is smaller than roughly 8 times the new input tokens for an Opus 5 to Haiku 4.5 switch, or when the turn produces a large output like a full file or long document, since output savings can offset the cold reprocessing cost. Short, independent prompts or the moment right after a context compaction (where the cache is already invalidated) are the cheapest points to switch. Developers designing multi-model agent pipelines can follow this kind of cost-routing analysis on daily.dev.