Performance / measured

Real latency, from a real client

Nobody should have to guess. These are client-side timings measured on 22 November 2026 from one machine in Asia, hitting the production endpoints with a marketplace key.

Endpointp50p95fastest
emailcheck /v1/verify172 ms554 ms147 ms
emailcheck /v1/bulk (10)163 ms185 ms156 ms
seo-audit /v1/audit (uncached)595 ms1359 ms267 ms
seo-audit /v1/audit (cached)211 ms227 ms199 ms
docparse /v1/html/text181 ms470 ms159 ms
docparse /v1/pdf/text 16p596 ms668 ms433 ms
textmind /v1/sentiment152 ms409 ms142 ms
textmind /v1/summarize2405 ms2939 ms2095 ms
pii-shield /v1/redact174 ms486 ms164 ms

Read the panel, not just the row

Every response carries its own server-side timing. Calling /v1/verify reports:

x-response-time-ms: 16
Server-Timing: cfEdge;dur=7, cfOrigin;dur=0, cfWorker;dur=21

So of the 172ms round trip, about 16ms was the Worker. The rest is distance: the request travelled from Asia to a Cloudflare edge location and back. A client in Frankfurt or Virginia will see a smaller number for the same endpoint, and no amount of code on our side changes that.

What the numbers show

  • The deterministic endpoints sit at the floor: ~150-180ms p50, which is essentially the network round trip, not compute.
  • The cached audit is 211ms p50 against 595ms uncached. That is the edge cache doing what the guide says it does.
  • A 16-page PDF takes about 600ms; the parse is the work, not the fetch.
  • The one neural endpoint measured, /v1/summarize, is 2.4s p50. That is a 20B model answering, and it is the honest cost of the feature.
  • p95 is two to three times p50 on several rows because these are single calls from one machine with no warmup, not a load test.

Method and limits

  • Node's fetch from one residential connection in Asia, 12 sequential calls per endpoint (4 for the model and PDF rows), no concurrency.
  • The audit rows use eight distinct URLs for "uncached" and ten repeats of one URL for "cached".
  • Timings include DNS, TLS and the full body read. They are end-to-end, not server-side only.
  • This is not a benchmark of peak throughput. It is what one caller experienced, which is the number a buyer actually cares about.

Re-run it yourself: node .tmp/bench.mjs against the live hosts. The script is in the public repo.