Understand your latency
Speed moves files. Latency is the wait for a reply. A useful connection needs both.
Our test compares HTTP replies before transfers, during downloads and during uploads. Every completed reading includes an explanation, with the evidence available under Show more data.
Read your three latency values
| Measurement | What it tells you |
|---|---|
| Before transfers | Your starting point on this HTTP path. Other applications can still be using the connection. |
| During download | Whether receiving data coincides with longer waits, as when a large download competes with interactive requests. |
| During upload | Whether sending data coincides with longer waits, as when cloud backup competes with a call or remote tool. |
The main values are p95: at least 95% of successful sampled replies took this long or less. With fewer than 20 successes, p95 is the maximum. The details include median, jitter, counts and failures. 100 ms is a tenth of a second; 1,000 ms is one second.
When does the result flag a problem?
| Observed evidence | Interpretation in this test |
|---|---|
| One or more failed requests | Some requests failed. Timings of successful requests exclude those failures. |
| Loaded p95 increases by at least 100 ms | Busy traffic coincided with added waiting. Investigate background transfers, the device link and router queues. |
| Any phase p95 reaches 250 ms | Replies were long on this path. A high baseline means the wait was present before test transfers. |
| No cue, with sufficient coverage | No large delay stood out in these samples. This is a short snapshot, not a guarantee. |
| Capped, short, sparse or incomplete | Keep the measurements, but there is not enough evidence to judge sustained behavior. |
The 100 ms increase and 250 ms absolute boundaries are our editorial investigation cues. They are not medical, call-quality, gaming or agent-productivity standards. The interpreter is iqm-interpretation-v3; the method page specifies the full rules.
Published reference points
Different tools use different samples, endpoints and statistics. These references explain the scale and the design; their scores are not directly interchangeable with our HTTP p95.
- FAST.com keeps its first result simple and reveals unloaded/loaded latency with additional information.
- Ookla's connection-quality guide explains latency measured at idle, during download and during upload, including the effect of competing transfers.
- Waveform's bufferbloat test publishes added-latency grading boundaries at 5, 30, 60, 200 and 400 ms. Its grading statistic differs from our directional p95 changes, so we do not assign its letter grades to these results.
- Cloudflare's measurement explanation describes throughput, loaded latency and why more bandwidth alone does not establish a responsive connection.
- Cloudflare Radar Internet Quality provides broader observed network context. Those measurements are not a local baseline or a percentile ranking for your reading.
What to try when waiting increases
Pause background backups or downloads. Compare Ethernet with Wi-Fi, or move closer to the access point. If you administer the router, investigate Smart Queue Management. Change one condition at a time and compare the same profile. No test runs automatically.
What this tells you about agents
Repeated network requests can accumulate waiting, but model processing, token generation, CPU and tools also take time. This test measures the host's HTTP path to Cloudflare. It does not measure the model provider or produce a validated productivity score. No controlled coding-task results are claimed here.