Compare / turbopuffer
WarpSearch vs turbopuffer
Search on object storage, like WarpSearch. Queries are billed by the size of the namespace they search, with a 1.28 GB minimum.
How they bill
| WarpSearch | turbopuffer | |
|---|---|---|
| Billing unit | Flat $2 per million requests | $1 per PB queried (namespace size per query, 1.28 GB minimum), discounts above 32 GB |
| Storage | $0.08 per GB-month, lower at volume | $0.33 per GB-month |
| Writes | $1.00 per GB | $2.00 per GB, up to 50% off for large batches |
| Data returned | $0.12 per GB, 100 GB a month free | $0.05 per GB |
| Minimum | $10 a month | $16 a month (Launch) |
| API | Elasticsearch, OpenSearch and Solr compatible | Own API |
| Hybrid and aggregations | Yes | BM25, server-side RRF, aggregations up to 10k groups |
Price
Monthly cost for four common workloads, at list prices in US East.
1M text chunks with 768-dim embeddings, 7 GB of JSON. 1M queries a month spread through the day. 1 GB written a month.
| WarpSearch | turbopuffer | |
|---|---|---|
| Storage | $0.56 (7 GB held x $0.08) | $2.31 |
| Writes | $1.00 (1 GB x $1.00) | $1 to $2 |
| Search | $2.00 (1M requests) | $7 (7 PB queried) |
| Data returned | $0 (about 40 GB, inside the free 100 GB) | $2 |
| Minimum | $10 minimum applies (usage is $3.56) | $16 applies |
| Total a month | $10 | $16 |
Usage is about $12 to $13.
5M products with 384-dim embeddings, 10 GB of JSON. 50M queries a month (about 19 a second on average). 20 GB written a month.
| WarpSearch | turbopuffer | |
|---|---|---|
| Storage | $0.60 (7.5 GB held x $0.08) | $3.30 |
| Writes | $20 (20 GB x $1.00) | $20 to $40 |
| Search | $100 (50M requests) | $500 (500 PB queried) |
| Data returned | $18 (250 GB returned, 100 GB free, before compression) | $12.50 |
| Minimum | Cleared | Cleared |
| Total a month | $139 | $536 to $556 |
10 TB of raw logs kept searchable. 1 TB written a month. 10,000 dashboard aggregation queries a month.
| WarpSearch | turbopuffer | |
|---|---|---|
| Total a month | $1,465 | Not priced |
Not built for log retention.
1 PB of raw logs kept searchable. 50 TB written a month. 100,000 heavy aggregation queries a month.
| WarpSearch | turbopuffer | |
|---|---|---|
| Total a month | $39,060 | Not priced |
Not built for log retention.
Performance
What turbopuffer publishes
| Vector, 10M x 768-dim, warm | 14 ms median, 27 ms p99 |
|---|---|
| Vector, same, cold | 874 ms median, 1,686 ms p99 |
| BM25, 1M docs, warm / cold | 13 ms / 316 ms median |
| Write (512 KB batch, durable) | 165 ms median |
| Writes per namespace | At most one commit per second; 10,000 writes/s |
turbopuffer publishes warm and cold numbers, measured from a client. Ours are measured next to the engine, so the warm numbers are not directly comparable. Our keyword median is 2 to 6 ms at 1 to 10M docs, and vector 6 to 16 ms at 0.95+ recall.
What we measured
| Test | Result |
|---|---|
| Keyword search, 2.8M docs (LoTTE) | 2.1 ms median, 3,700 queries/s |
| Keyword search, 10M passages (MS MARCO v2) | 2.4 ms median (5.4 ms with an exact total hit count) |
| Keyword search, small BEIR sets | 1 to 6 ms median |
| Vector search, 10M x 128-dim (BigANN SIFT) | 0.964 recall@10 at 6.2 ms median, 162 queries/s |
| Vector search, 1M x 1536-dim (arXiv) | 0.961 recall@10 at 15.7 ms median |
| Filtered vector search, 10M (YFCC) | 0.959 recall@10 at 10.4 ms median |
| Log filters, 22M events (http_logs) | 1 to 2 ms median |
| Log aggregations, 22M events | 190 to 430 ms median (sort, daily histogram, avg by status) |
| Ingest, short text | 63,000 to 97,000 docs/s |
| Ingest, logs (247M events) | 76,000 to 81,000 events/s |
| Ingest, 768 to 1536-dim vectors | 2,200 to 4,000 vectors/s |
| Time to searchable | Seconds, or immediate with commit=true |
| Cold start | Not measured yet. First query on a 2.3 GB index: 17 to 19 ms vs 2.7 ms warm (page cache only) |
Measured October 2026 through our Elasticsearch-compatible API, warm cache, client in the same region, single runs. Every test ran on deliberately minimal hardware, 2 to 4 CPUs (a 4 CPU container or an AWS i8g.large with 2 vCPUs), so these are the least favorable results we can produce. More hardware only improves them. Numbers exclude internet round trips.
Bottom line
Choose WarpSearch for
- Query-heavy workloads: our request price does not grow with index size. On the catalog workload we are about 4x cheaper.
- Elasticsearch, OpenSearch and Solr compatibility: keep your clients and queries.
- Storage at a quarter of the price per GB.
Choose turbopuffer for
- Published cold-start latency. We have not published ours yet.
- Proven scale: turbopuffer publishes results at 100B vectors.
Same architecture, different meter. WarpSearch charges per request; turbopuffer charges per GB searched. The more you query a large index, the more WarpSearch saves. For a small, rarely queried index, the two cost about the same.
Sources
The $0.05 per GB returned rate comes from turbopuffer's pricing calculator; the FAQ only calls it flat. Prices checked October 20, 2026. List prices in US dollars, before discounts. Estimates are marked. Tell us if something is out of date: contact us.