Turbopuffer
Vector + BM25 search on S3 — cheap when data goes cold
Overview
Turbopuffer indexes vectors and full-text on object storage rather than RAM, undercutting RAM-resident databases like Pinecone by 10x or more at scale. Cursor, Notion and Anthropic run production search on it.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- Pay per GB stored + per PB queried
- Batch write discounts up to 50%
- No free tier
- HIPAA BAA
- SSO
- Audit logs
- Single-tenancy
- Bring-your-own-cloud (BYOC)
- Customer-managed encryption keys (CMEK)
- Private networking
Use cases
What you can produce with Turbopuffer
- Hybrid BM25 + vector search in a single query
- Namespaces scale to millions with no per-namespace overhead
- Sub-10ms warm-cache query latency
- Official Python, TypeScript, Go and Java clients
- Runs on S3, GCS or Azure Blob storage
- HIPAA BAA, SSO and audit logs on the Scale plan
- BYOC, CMEK and single-tenancy on Enterprise
ASEAN Perspective
Turbopuffer in Southeast Asia
ASEAN-region availability and pricing notes coming soon. Drop the editorial team a note via /contact/ if you can supply local context (Singapore/Malaysia/Indonesia/Thailand/Vietnam).
Turbopuffer's bet is that most vectors sit idle between queries, so keeping them in RAM at all times is a waste of money. It stores indexes on S3 (or GCS/Azure) with a hot cache for whatever got hit recently, and that architecture is why Notion could drop its per-user AI storage surcharge after migrating, and why Cursor now indexes over a trillion code chunks across 80 million-plus namespaces on it.
The trade-off shows up on cold namespaces: the first query after a quiet spell pays a latency tax while data loads back from object storage, so anything needing flat, guaranteed low latency on every single query should look elsewhere. It's also a 2023 company with a much thinner ecosystem than Pinecone or pgvector, and there's no free tier — the cheapest plan floors at $16/month regardless of usage. Worth it when your storage bill is the actual pain point.
What people say
Simon Hørup Eskildsen and Justin Li, both ex-Shopify infrastructure engineers, started turbopuffer in 2023 with a pitch that sounded almost too simple for Hacker News: "$1/month per million vectors" by building on object storage instead of RAM. The original HN threads split between people impressed by the cost math and people skeptical that cold-storage latency would hold up in production — three years on, the latency argument looks mostly settled. The company stayed close to profitable on under $1 million in total funding before taking an undisclosed seed round from Lachy Groom and Thrive Capital in December 2025, and hit $100 million in annualized revenue by March 2026, up from $75 million at the end of 2025.
The customer list does the talking: Cursor uses it to index more than a trillion code chunks across 80 million-plus namespaces, and Notion's migration of its 10 billion-plus vector workload let the company drop per-user AI storage charges outright. Anthropic, Atlassian, Linear, Grammarly and Clay are also named production users.
Pricing has moved in the buyer's favor twice this year — the per-PB queried-data rate dropped from $5 to $1 in February 2026, and the Launch tier's monthly minimum was cut from $64 to $16, making it cheaper to try at small scale than it used to be. Storage itself is the number engineers cite most: roughly $0.02/GB against Pinecone Serverless's $0.33/GB, which is why teams running 100M-1B vector workloads report $500-$2,000/month bills against $5,000-$20,000/month equivalents elsewhere.
The recurring criticism, on HN and in independent write-ups, is the one the architecture invites: cold-query latency. First-touch queries on data that hasn't been accessed recently take roughly 300ms while it loads from object storage, before subsequent queries drop to sub-10ms from cache — fine for idle-heavy workloads like code search, less fine if every query needs to be fast regardless of access pattern. There's also no self-serve free tier, so evaluating it means either talking to sales or eating the $16/month floor.
Summary of public user & expert reviews, compiled by RECATOOLS.
About this listing
This entry was compiled from publicly available data including Turbopuffer's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Turbopuffer unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Turbopuffer directly →
Spotted something out of date? Suggest an update →
More in Code & Dev Tools
