KV cache — why the second token is cheaper than the first
The KV cache stores attention keys and values so each new token does not recompute the whole context, which is why prefill is expensive and decoding is fast — and why cache pricing reshapes agent economics.
Not yet scoped.
Why
The PoC measures prefill versus decode cost with and without cache reuse, connecting the mechanism to the cache-read price cut that changes long-context agent bills.
How it works
Not yet built.
Where it lands in Jayverse
Auditor / gitboard: track prefill vs decode cost separately for any LLM-backed tool. Agent workflows and Auditor methodology write-ups should prefer cache-read pricing for repeated long-context calls, since that's where the real bill lives.
Devnet / alice-tech: reuse the same context prefix across workflow stages. Running agent-team-style stages against a fixed case record only gets cache reuse if each stage doesn't re-pay prefill from scratch.
Key expressions
Words and phrases from this page worth keeping, with the Korean meaning and the sentence they come from.
Expression
뜻 · 쓰이는 자리
reshape
(구조·형태를) 다시 짜다, 재편하다 · 가격 정책이 전체 비용 구조를 바꿀 때. "why cache pricing reshapes agent economics"
recompute
다시 계산하다 · 이전에 계산해둔 것을 매번 새로 구하지 않아도 될 때. "does not recompute the whole context"
versus
~ 대(對) ~, ~와 비교하여 · 두 단계나 항목을 나란히 비교할 때. "measures prefill versus decode cost"
cache reuse
캐시 재사용 · 이전에 계산한 결과를 다시 활용하는 것. "with and without cache reuse"
price cut
가격 인하 · 특정 항목의 가격이 낮아진 것을 가리킬 때. "the cache-read price cut that changes long-context agent bills"
KV cache
키/값 캐시(Key-Value cache) · 어텐션의 키·값을 저장해 두어 새 토큰마다 전체 문맥을 다시 계산하지 않게 하는 캐시. "The KV cache stores attention keys and values"
prefill / decode
프리필(프롬프트 전체를 한번에 처리하는 단계) / 디코드(토큰을 한 개씩 생성하는 단계) · LLM 추론이 나뉘는 두 단계, 비용 구조가 서로 다름. "why prefill is expensive and decoding is fast"