Cache

$0.0033 saved

Three cost levers: exact-response reuse, singleflight coalescing of identical in-flight requests, and provider prompt-cache affinity that steers repeat prefixes back to warm placements.

Exact hits

6rate 25%

Misses

188 stored

Coalesced

0singleflight

Bypass

18ineligible

Saved · exact

$0.0033responses reused

Saved · prompt cache

$0.00discounted input

Exact-cache entries

Post-guardrail responses keyed by canonical request hash. Guardrail-policy and profile versions are part of the key, so a policy bump invalidates without a flush.

Key prefix	Model	Keyed versions	Tokens	Age	TTL left	Hits
9255a633d03395…	anthropic/claude-sonnet	gp_baseline@3@company/general-balanced@7	2.0K	4m ago	56m	2
2d741b0f9e3f85…	openai/gpt-5-mini	gp_baseline@3@company/general-fast@4	516	22m ago	4m	7

Prompt-affinity directory

Stable prompt prefixes fingerprinted and mapped to the provider/account/node where the provider-side prompt cache is warm. The router pays an affinity bonus to keep repeat prefixes on these placements.

Prefix fingerprint	Model	Provider / account / node	Cached tokens	Confidence	Expires
pfx_7f3a91c2	anthropic/claude-sonnet	anthropic · acct_anthropic_ent_apse1 · node_sg_dc_01	12K	84%	4m
pfx_c04d22ee	openai/gpt-5.2	openai · acct_openai_pool_b_apse1 · node_sg_dc_01	6.1K	61%	2m

Session directory

Active sessions pinned to a placement so multi-turn conversations keep hitting the same warm cache.

Session	Pinned model	Endpoint / node	Cached tokens	Confidence	Expires
sess_agent_ops_01	anthropic/claude-sonnet	ep_sonnet_anthropic_apse1 · node_sg_dc_01	18K	92%	5m

Cache invariants

· Input guardrails run before any cache lookup — a blocked request never reads the cache.
· Only output-approved content is written; the write happens after POST_MODEL/PRE_RETURN pass.
· Guardrail-policy and profile versions live inside the cache key, so stale-policy responses can never be served.
· Confidential profiles run with mode: off — nothing is stored.