OrcaCyber K3.1 Pro: A Week That Improves Every Day and Caps Twice
Eleven days of consecutive improvement is not what a serving week usually looks like, and this card managed it: a p50 that fell from 1,406 milliseconds to 875 across seven days. OrcaRouter reports the daily series rather than only the aggregate. The series is on OrcaCyber K3.1 Pro latency, and it is the most steadily improving week in the tier.
The seven days
| date | p50 ms | p95 ms |
| 2026-11-14 | 1,406 | 4,467 |
| 2026-11-15 | 1,421 | 4,859 |
| 2026-11-16 | 1,599 | 10,000 |
| 2026-11-17 | 1,470 | 7,337 |
| 2026-11-18 | 1,046 | 8,435 |
| 2026-11-19 | 994 | 10,000 |
| 2026-11-20 | 875 | 4,827 |
The weekly p50 of 1,069 ms is a blend of a median that started at 1,406 and ended at 875 — a 38% improvement over six days. That is a real trend rather than noise: three of the four middle days sit between 1,421 and 1,599, then the last three fall to 1,046, 994 and 875.
Monotonic improvement in a serving median usually means one of three things: a routing change, added capacity, or a workload shift toward shorter requests. The published fields cannot distinguish them.
The p95 column is where the week gets interesting. Four days carry a readable tail — 4,467, 4,859, 7,337, 4,827 — and three days hit the 10,000 ms cap. The two capped days, the 16th and the 19th, are not consecutive, so this is not one incident.


What a capped p95 actually tells you
A p95 of exactly 10,000 ms means the measurement window closed rather than that the 95th percentile was 10 seconds. The true tail on those days is unknown and at least 10,000.
That matters for timeout design. A client that sets its timeout from the readable days — say 5,000 ms — would have been correct on two days and wrong on three. A client that sets 10,000 ms would clip the unmeasured tail on three days and wait needlessly on two.
| card | spread (max p50 ÷ min p50) | capped days |
| Hy4 | 1.09× | 3 |
| MiniMax M3.1 Flash | 1.20× | 1 |
| GLM 5.3 Flash | 1.34× | 7 |
| GLM 5.3 | 1.41× | 1 |
| Flash 1.0 | 1.57× | 0 |
| Hy4 Flash | 1.33× | 1 |
| K3.1 Pro | 1.83× | 2 |
| Zero 1.5 | 2.26× | 4 |
| MiMo V2.6 Flash | 2.53× | 6 |
| Zero 1.0 | 3.08× | 7 |
| Ultra 1.0 | 3.37× | 6 |
This card’s 1.83× spread sits in the middle of the tier, and its two capped days are fewer than five other cards manage. So although the week improves, its variability is not unusual — the improvement is a level shift rather than a stabilisation.
Fast first token, slow generation
The card’s throughput is 81.4 output tokens a second, eleventh of twelve. Its p50 time to first token is 1,069 ms, second-fastest of twelve.
That combination is distinctive. Most cards move both figures together: Flash 1.0 is fast at both (940 ms, 210.9 tok/s), MiMo V2.6 Flash is slow at both (5,757 ms, 42.3 tok/s). This card is fast to start and slow to finish.
| card | p50 ms | tok/s | implied median tokens |
| Flash 1.0 | 940 | 210.9 | 198 |
| K3.1 Pro | 1,069 | 81.4 | 87 |
| Zero 1.5 | 1,394 | 195.6 | 273 |
| MiniMax M3.1 Flash | 1,691 | 110.3 | 187 |
| Qwen3.8 Flash Next | 1,821 | 91.0 | 166 |
| Zero 1.0 | 1,883 | 240.2 | 452 |
| Ultra 1.0 | 2,970 | 109.2 | 324 |
| GLM 5.3 | 3,411 | 91.6 | 312 |
| Hy4 Flash | 3,579 | 182.1 | 652 |
| Hy4 | 3,599 | 135.1 | 486 |
| GLM 5.3 Flash | 4,144 | 66.4 | 275 |
| MiMo V2.6 Flash | 5,757 | 42.3 | 244 |
The last column multiplies the median by the rate, and it is our arithmetic rather than a published field. At 87 implied tokens this card’s typical answer is the shortest in the tier — a quarter of Flash 1.0’s 198 and a seventh of Hy4 Flash’s 652.
So the fast first token is not compensating for a long answer. The card starts quickly and produces a short answer, slowly. That is a coherent profile for a model that reads a large prefix and returns a verdict.
Reading the week as a buyer
Three facts to plan against, all from the published series.
The level moved. The last three days sit between 875 and 994 ms. If that level holds, the card’s weekly p50 of 1,069 overstates its current latency by about 20%. If it does not, the earlier 1,400–1,600 band is the fallback.
The tail is unmeasured on two days in seven. Building a timeout from the readable days alone under-provisions by an unknown amount.
The p95 and the p50 do not move together. On the 14th the median was the week’s highest and the tail was the week’s tightest at 4,467. On the 16th the median was high and the tail capped. So a slow day is not necessarily a wide day here, which is the opposite of what Zero 1.5 shows.
That last point is the one worth carrying into a monitoring design: a threshold tuned to this card’s p50 would flag the wrong days, because the p50 and the p95 are not measuring the same population.
Where a 38% weekly improvement could come from
Three mechanisms produce a falling serving median, and the published fields rule none of them out.
Capacity added. More workers behind the same endpoint shortens queue time, which shows up in the first token and not in the generation rate. The rate is unchanged at 81.4 tokens a second, so the improvement is consistent with this.
Routing changed. If the endpoint began preferring a warmer or nearer pool, the first token would fall without any change in throughput.
Workload shifted. Shorter inputs mean a shorter prefix read, which lowers the first-token time. The implied median answer also fell across the week, which is consistent with smaller requests.
| date | p50 | implied tokens | p95 |
| 11-14 | 1,406 | 114 | 4,467 |
| 11-17 | 1,470 | 120 | 7,337 |
| 11-18 | 1,046 | 85 | 8,435 |
| 11-20 | 875 | 71 | 4,827 |
The implied column is our arithmetic, and its decline tracks the median exactly because the rate is constant. So either the answers got shorter or the queue got shorter — and those two would look identical in this table.
A buyer cannot distinguish them and does not need to. The figure to carry forward is that the card reached 875 ms and that the same card was at 1,406 six days earlier, so any capacity plan should assume a range rather than a point.
There is a fourth candidate the published fields cannot address: the definition of the median itself. If the probe’s request mix changed mid-week — more short health checks, fewer long generations — the median would fall without any change in serving. That possibility is mentioned for completeness rather than as a claim, and it is the reason the article stops at describing the trend rather than attributing it.
OrcaCyber K3.1 Pro’s seven daily medians fell from 1,406 to 875 milliseconds — a 38% improvement across the week and the steepest sustained drop in the tier — while its p95 hit the 10,000 ms cap on two of seven days.
Its 1,069 ms first token is second-fastest of twelve cards, but at 81.4 tokens a second the implied median answer is about 87 tokens, the shortest in the tier. Fast to start, short, and slow to finish is a verdict-shaped profile, and the unmeasured tail on two days is the planning risk.
Sourcing note: the seven-day daily p50 and p95 series, the weekly percentiles, throughput, error rate and price rows are read from the OrcaCyber K3.1 Pro model API and card page on orcarouter.ai on 2026-11-20, with latency and throughput being OrcaRouter production telemetry for 2026-11-14 to 2026-11-20. The spreads, implied token counts and comparisons are our own arithmetic. A p95 reading of exactly 10,000 ms is the reporting cap, not a measurement. No competitor product is named or compared. Checked 2026-11-20.
