EDIT :
- solving auto-caching from Anthropic by updating ai-sdk/anthropic +
adding providerOption at stream level
- concerning Bedrock, it needs breakpoint
**1. Breakpoints were only on the system prompt**
The code already placed a cache marker on the system prompt (~10K
tokens). But the conversation history — which can grow to hundreds of
thousands of tokens — had no marker, so Anthropic re-read it at full
price on every turn.
The fix adds a prepareStep hook inside streamText that stamps the last
message with a cache breakpoint before every LLM call. Anthropic then
caches the entire conversation prefix, and subsequent turns read it at
$0.30/M instead of $3/M.
prepareStep is used rather than a one-shot pre-processing step because
an agentic turn makes multiple internal LLM calls as tool results
accumulate — the hook refreshes the breakpoint before each one.
**2. Bedrock was using the wrong field**
The system prompt marker for Bedrock was set as cacheControl: { type:
'ephemeral' } — which is the Anthropic wire format. The Bedrock Converse
API expects cachePoint: { type: 'default' }. The system prompt was
silently not being cached on Bedrock at all.
Both the system prompt and the new prepareStep now go through a shared
getCacheProviderOptions helper that returns the correct field per
provider.
**3. Persisted cached token usage to monitor cache strat. efficiency**
