Benchmarks, measured honestly

Everything on this page was measured with VectorDBBench — the open-source benchmark harness maintained by Zilliz — running against the same public API every customer uses. No internal shortcuts, no synthetic data, no hand-picked runs. Raw harness output is downloadable below, including the run our own rate limiter spoiled.

Last verified 2026-07-21 · VectorDBBench 1.0.22
1,620
QPS, sustained
0.965
Recall@100
<50ms
Serial p99
$17.14
This collection, per month
The setup

Standard dataset, public path

VectorDBBench's Cohere 1M case: one million 768-dimension embeddings, cosine distance, recall@100 scored against the dataset's own ground truth. The index was chosen by our automatic optimizer — nobody hand-tuned anything.

DatasetCohere 1M (768d), VectorDBBench standard case, cosine, k=100
Path measuredPublic API (api.vectorpanda.com) — TLS, auth, the works
IndexAuto-selected by the optimizer (HNSW family); zero manual tuning
Recall@1000.9652 (NDCG 0.9667), scored on the dataset's full ground truth
Serial latency p9949.5 ms
Max sustained QPS1,620 (flat from 100 to 200 concurrent clients)
Errors during the sweep0 across ~236,000 queries
Collection price$17.14/month (2.86 GB hot × $5.99/GB-month, list price — queries included)

Raw harness output: this run (JSON) · the rate-limited first attempt (JSON)

The curve

Throughput by concurrency

30 seconds per step, VectorDBBench's multiprocess runner. Throughput reaches its plateau around 100 concurrent clients and holds flat while latency grows linearly — a saturated system queueing politely, with zero failures.

Concurrent clientsQPSp95p99
14642 ms51 ms
1048624 ms53 ms
2087228 ms59 ms
501,46241 ms83 ms
1001,61988 ms138 ms
1501,620156 ms198 ms
2001,604179 ms247 ms
The run we spoiled ourselves. Our first full-path attempt measured 1,434 QPS, declining above 50 clients — which turned out to be our own per-key rate limiter (since raised), not the database. We publish that run's data too: comparing the two tables is a compact lesson in how easily a benchmark measures the wrong ceiling, and why raw output beats a headline.
Context

Against published VectorDBBench results

Rows below are from VectorDBBench's published July 2026 results for the same Cohere 1M case; monthly prices are each vendor's list price for the configuration benchmarked. Ours is the list price of the collection itself.

SystemQPSRecallp99$/monthQPS per $/mo
ZillizCloud 8cu-perf (best row)12,8380.95882.1 ms$92913.8
OpenSearch 16c128g2,5900.96798.6 ms$1,0352.5
Vector Panda1,6200.965249.5 ms$17.1494.5
ElasticCloud 8c60g1,4820.954612.1 ms$9201.6
Pinecone p2.x81,1460.926213.7 ms$9561.2
QdrantCloud 16c64g1,1110.95507.0 ms$8321.3
TurboPuffer (serverless)7980.899356.7 ms
S3Vectors (serverless)1990.8717337 ms

Comparisons only make sense inside a recall band: our 0.9652 sits between Zilliz's best row (0.9588) and OpenSearch (0.9679). Vendors' numbers come from their published VectorDBBench runs on their hardware; we have not re-run their systems ourselves.

Read it honestly. Zilliz and OpenSearch beat us on absolute throughput — if a single collection needs more than ~2,500 QPS today, or you need single-digit-millisecond p99, those are the right tools and they are priced for it. Where the comparison tilts hard the other way is throughput per dollar: this collection serves 1,620 QPS for $17.14 a month — about 94 QPS per dollar-month against 13.8 for the best published row above. Queries are included in the price, so the ratio doesn't decay with traffic.

Reproduce it

Run it yourself

The harness is open source, the dataset is standard, and the API is the public one. A skeptical afternoon is all it takes.

  1. Install VectorDBBench (pip install vectordb-bench) — it downloads the Cohere 1M dataset and ground truth itself.
  2. Grab our VectorDBBench client and runner from the veep repository — it wraps the veep SDK's create/upsert/query calls, nothing more.
  3. Point it at your account with --prod, load the dataset, let the optimizer converge, and run the sweep. The collection you create bills like any other (about $17/month hot for this dataset — delete it when you're done).
  4. Compare your JSON to ours. If your numbers disagree with this page, we want to know: tell us.