All posts
Product#Product#Launch#Caching
Introducing Semantic Cache v2 — now with embedding-aware invalidation
Our second-generation semantic cache lands today: embedding-aware invalidation, per-tenant namespacing, and 38% fewer input tokens on average.
TB
Theo Bauer
Product, SpendTensor
May 21, 2026 5 min read
Introducing
Semantic Cache v1 shipped 14 months ago and now serves 2.1 billion requests a month. v2 is a from-the-ground-up rebuild based on what we learned.
The headline change is embedding-aware invalidation. When upstream context changes, we invalidate only the cache entries whose embeddings are semantically affected — not the whole namespace. Across launch customers we're seeing a 38% average reduction in input tokens with zero observed staleness regressions.
v2 is generally available today on Growth and Enterprise plans. No code changes required.