Your speed test measured an idle line at full throttle and told you 480 Mbps. The thing that ruined your gunfight lasted a second and a half, happened while somebody was uploading, and consisted of packets that turned up after they were any use. Bullettime is built to catch exactly that.
No account, no card The free result is the whole result Runs on your own box for $9.99
Warzone is likely degraded while anyone is uploading.
Direct pathThe hole is the headline because it is the figure that survived our own testing (we published the workings). Underneath it, the composition, never merged into one number: while the line was busy going up, the 413 probes in that epoch sorted as:
n = 413 95% CI ±4.4 pp
Everything left of the line made it. Everything right of it was delivered, and was useless.
The idea the whole product is built on
A 200 ms arrival is fine for Netflix, survivable for a call, and gone for a 60 Hz shooter, so the honest thing is to refuse to pick your deadline for you and give you the axis instead. Drag it. How much a packet’s lateness costs you is the part we have actually measured; how best to summarise it in one number is the part we got wrong first time.
Deadline
16.6 ms
4 ms240 ms
On time 0
Arrived too late 0
Never arrived 0
Deadline‑miss rate 0%
n = 413upload epoch of the sample session
This scene needs WebGL, which this browser has turned off.
The same figures are in the panel beside it, and the deadline control still works.
One volley, frozen: every probe sent during the eighteen seconds the line was busy going up, placed by how long it actually took. Hollow marks past the plane arrived and were useless. Solid marks never arrived at all. The scene is a picture of a distribution rather than a decoration, and moving the plane runs the same arithmetic the product runs.
Why a percentage is the wrong shape
Loss clusters. Measured traces run 20–30% density inside a burst for several seconds at a time, and a mean percentage cannot tell the two pictures below apart. So Bullettime keeps the raw received‑or‑lost sequence for every probe and reports the shape: run lengths, the longest hole in milliseconds, and how many holes per minute.
longest hole 21 ms2 holes / min
Invisible. Concealment covers it and nobody notices.
longest hole 1,420 ms1 hole / min
A frozen face, a dropped call, a gunfight you were never in.
This is the best-evidenced claim on the page. We put the longest hole up against conventional packet loss, mean latency, RFC 3550 jitter and RFC 5481 delay variation across 1,344 simulated sessions, pre‑registered, on a fresh seed, and it beat all of them. It also beat the deadline‑miss percentage we used to lead with, which is why we no longer do. The workings are published, including the two runs that went against us.
Microsoft’s own call-quality documentation makes the same argument: a 1.5 second gap dilutes into a low whole‑call average while producing 30% loss inside a five‑second window. Every other consumer tool reports the average.
The measurement contract
Fill the buffers, measure latency during it, then call the result “your latency”, and you have reported the test’s own effect as a property of the network. So a session is a sequence of epochs, each with its own low‑rate probe lane on a separate flow and a separate server queue. Loaded figures are reported as loaded figures and never merged with idle ones.
The clock below is the point. Pick any epoch and it jumps, because these are sequential minutes apart on purpose. Nothing here is one simultaneous instant, and a tool that renders it as one is showing you something it did not measure.
Session clock at this epoch
00:04 → 01:02 the only epoch allowed to be called idle
Loading…
What you actually get back
These are the real surfaces, not illustrations of them. Nothing below is withheld from the free tier to manufacture a reason to upgrade; what the paid plans buy is the work a browser tab structurally cannot do.
Deadline survival curve — misses against assumed buffer depth
n = 41395% CI ±0 ppT obs = 3,000 ms
A clean result never prints “0%”. It prints the measured figure with the sample count and interval that make it a claim rather than a reassurance.
Received or lost, per probe, across the loaded epochs
A Gilbert–Elliott fit is available and is suppressed here, because this trace holds too few state transitions for the parameters to be stable. Fitting a two‑state model to a nearly lossless trace invents structure.
Call of Duty: Warzone
Likely degraded now
Zoom
At risk when busy
WebRTC audio
At risk when busy
Netflix
Likely fine
GeForce NOW
Path mismatch
Five verdicts, and one of them is “we cannot assess this from here”. Every competitor ships per‑application presets with no provenance at all.
288 windowslight probe, 5 pps, 10 s, every 5 min
Your line got worse at 20:05, and it is still worse.
Deadline misses have run six times your own weekday evening baseline for 47 minutes. Calls and games are the things this affects. It is not a single bad sample.
Fired against this household’s own established baseline, not a threshold. Why did this fire?
A period with too little data is reported as unmonitored, never as healthy.
PDF p.2PDF p.3PDF p.4PDF p.5JSONmanifestWritten to be read by someone who did not run it, which means it carries its own methodology rather than assuming the reader knows what the columns mean. Re‑exporting the same range produces the same document.
Available on the free tier for a single result, because arguing with an ISP is the free tier’s best story and gating it would be the wrong side of our own rule.
every figure on this page derives from one seeded sample session
Product rules, not aspirations
Each one exists because a tool you have already used prints it, and each one misleads. The bars below do not lift on hover. That is the point: the claim was never made.
A browser‑only remote test cannot separate your Wi‑Fi from your ISP, and saying so sends people out to buy routers they do not need. Separating them needs concurrent probes to your own gateway, which is what the native apps do and what a tab cannot.
Routers forward traffic in hardware and generate the ICMP that traceroute reads on a rate‑limited, deprioritised path. Perfectly healthy routes routinely display alarming mid‑path loss. Any path view we ship is labelled route observations, not fault locations.
Zero out of forty probes means your real loss could be anything up to about 8.8%. A figure without its sample count and confidence interval is a reassurance rather than a measurement, so ours never ships without both.
That was the round trip to our server, in our region. It is good evidence about your access link and no evidence at all about the path to Activision. When the endpoint does not represent the application, the verdict says path mismatch instead of inventing a number.
iOS does not permit it. Background refresh is discretionary and historically gets about thirty seconds per launch, which is an event detector and not a monitor. No plan on our pricing page claims it, and the pricing page says so where a buyer would look rather than in a footnote.
The free tier carries advertising, and networking hardware, ISPs and carriers, VPNs and “ping boosters” are excluded categories, published rather than merely honoured. A router advert beside a result reading this is consistent with queueing on your Wi‑Fi delivers, through a third party, exactly the conflict this product exists to avoid.
Five platforms, one measurement core
Same state machine, same statistics, same impact engine, same versioned result on every one of them. What differs is what each platform will actually let a program do, so each card says which rather than pretending the set is uniform.
Native app + service
Tray presence, monitoring that survives a reboot, raw sockets with TTL control, per‑hop history and Native Wi‑Fi telemetry.
Runs the gateway‑versus‑remote comparison, which is the only honest way to separate your Wi‑Fi from the line.
Native app + launchd agent
Menu‑bar condition and the age of the last sample, CoreWLAN RSSI and PHY rate, ICMP probes, thirteen designed surfaces.
Monitoring stopped, paused and running‑clean are three visibly distinct states, because a silent status area otherwise reads as a healthy connection.
Native app + Live Activity
Sixteen surfaces, a read‑only Live Activity while a test runs, and a Lock Screen that pins the dark palette rather than inheriting one that flips.
No background monitoring, here or on any plan. iOS does not permit it, whatever a competitor’s marketing says.
Native app, multi‑column
The sidebar build of the same thirteen surfaces, sized for a pointer and a keyboard rather than a phone scale stretched to fill.
Same limit as iPhone on background work, and the pricing page states it rather than burying it.
Headless agent + probe server
The reference agent for a NAS, a homelab box or an OpenWrt router, plus the probe server itself. Configured and read without a display attached.
A monitor on a laptop measures the laptop: it sleeps, it moves rooms, it joins a different network. This is the one that does not.
The web test runs in any modern browser and needs nothing installed. The native clients exist for the two things a browser genuinely cannot do: run when it is closed, and probe your own gateway.
Run the whole thing yourself
A measurement tool that publishes its server is making a checkable claim about its method. The probe server, the measurement spec and the result schema are open; any client can be pointed at your endpoint and complete a full session against it.
Your household’s connection history reveals your occupancy patterns. That is a large part of why local‑first and self‑hosted modes exist at all.
Two of the platforms people reach for first cannot accept inbound UDP, so a self‑hoster who tries either will hit a failure that looks exactly like a bug in this software. Saying so costs one paragraph and saves the support load.
| Target | Inbound UDP | TURN relay | Verdict |
|---|---|---|---|
| A $5 VPS Hetzner, Linode, EC2, GCE | yes | yes | The documented self‑host target. One container, one command. |
| NAS or homelab box | yes | behind NAT | Works, and gives you the always‑on wired reference that the household comparison needs. |
| OpenWrt / pfSense router | yes | no | The agent cross‑compiles for MIPS, ARM64 and x86_64. At the gateway there is no Wi‑Fi contention in the way, so it is the purest view of the line. |
| Fly.io | restricted | no | Conditional, and we will not promise it until a deployment test settles it. Needs a dedicated IPv4, binds fly-global-services, has no UDP load balancing, and cannot open the wide dynamic port range TURN allocation needs. |
| Vercel | no | no | Website, control API and result viewer only. There is no inbound UDP listener to bind, and there will not be one. |
| Cloudflare Workers | no | no | Same story. Excellent for the web tier, structurally unable to host the probe. |
Licence: probe server and shared analysis core under MPL‑2.0, which permits private and commercial self‑hosting and asks only that modifications to covered files stay available. Measurement spec and result schema under Apache‑2.0.
Pricing
The free test carries the same deadline‑miss figure, the same survival curve, the same burst view, the same path provenance and the same export as the paid one. What you pay for is the work a browser tab structurally cannot do: running when it is closed, telling you, and probing your own gateway.
Hosted by us
$0forever
One test at a time, in a browser, on whatever you happen to be sitting at.
no account, no card
Hosted by us
$4.99/ month
Round‑the‑clock monitoring of one address, because the faults that matter are the ones that are not happening when you go looking.
cancel any time, keeps exporting for 30 days
Background monitoring on iPhone and iPad. iOS does not permit it, so this plan does
not claim it. Monitoring runs from a Mac, a Windows machine, or a Linux box that stays on.
Hosted by you
$9.99once, not a subscription
Run the probe server on your own VPS or your own hardware, and keep the whole history on a machine you own.
one payment, all future versions
Bandwidth and abuse controls are on by default rather than opt‑in, because a saturating test at 1 Gbps moves about 3.75 GB and an open one will get found.
Watch and Self‑hosted are not tiers of each other. One is us running the probe and telling you when your line changes; the other is you running the probe and nothing leaving your network. Plenty of people want the second and nothing else.
What we refuse to sellWe tested our own claim. Part of it broke. 26 August 2026
The mechanism is not in question and never was. ITU‑T G.107 has folded late arrivals into effective loss since 2015; RFC 3611 and RFC 7002 define discard metrics separately from network loss; G.1072 and P.1203 treat a frame past its deadline as lost. A packet that arrives too late really is gone, and the standards have said so for years.
We built a game‑client simulation whose only measure of harm is how far, in metres, the client drew an opponent from where they actually were. Measured frame by frame, the mechanism is close to deterministic: the correlation between how stale the client’s information was and how wrong its picture was came out at 0.95, and the damage climbs from a millimetre at 25 ms of staleness to almost two metres past 200 ms. Late packets wreck netcode, and the harm scales with exactly how late.
What broke was our instrument, not the idea. A miss percentage is a count: it scores a packet 2 ms past the tick the same as one 400 ms past it, and throws away the magnitude that does the damage. Pre‑registered and run three times, that percentage lost to plain packet loss every time. It is the worst‑performing form of the metric we have measured.
What won, confirmed on a fresh seed, is the figure this page already puts beside the two grids above: the longest continuous hole, in milliseconds. How long you went without current information, whether the packets were late or never turned up at all. It beat conventional loss by a clear margin on both client models we tried, and it beat our own headline percentage by a mile. So the percentage has come off the front of this page and the hole length has gone on it. Scroll back up; that is what the headline figure is now. We have not commissioned the human study; the winning number has changed twice this week and should settle first.
Read the study and the data The preregistrations
The preregistrations, the results and the raw JSON are published whether they suit us or not. Provenance marking used throughout the product: a figure is either checked against the primary document, asserted by research we have not independently verified, or nobody has established it. A number with no mark is a design parameter we chose, not a finding. The figures elsewhere on this page come from a seeded sample session and not from a real connection.