On OrcaRouter the same card has been measured twice in four days. Four days is a short time in a model’s life, and long enough for its serving profile to move by a factor. OrcaCyber Zero 1.5 read a p50 of 1,875 ms and 116.2 tokens per second in the 2026-11-02 snapshot on OrcaRouter. The 2026-11-06 snapshot reads 3,553 ms and 51.3 tokens per second [OURS]. The price — $3 in / $7.50 out / $0.75 cache — is identical. So is the context, the ceiling and the access gate.
That is a 89% higher median and 56% less throughput for the same money, and it is the kind of change that never appears in a launch post. This piece is about what moved, what did not, and how to price a card whose shape is not stable.
What stayed the same
Everything a buyer reads on the page before they look at the timing column.
Price. $3 in, $7.50 out, cache at $0.75 — a quarter of the input rate, the steepest cache ratio in the security line.
Shape. 1,000,000-token context, 128,000-token output ceiling, text in and text out, tools, JSON and reasoning.
Access. `sec-uncensored` tier, `offensive` band: an engagement, a passkey and accepted terms before the first request.
Benchmarks. CVE-Bench 95.8 on the evaluable subset, Cybench 100 under unrestricted agent execution, HumanEval+ 93.9, SWE-bench Pro V2 76.5 — the vendor’s own rows, unchanged and unreproduced [VENDOR].
Those are the numbers a procurement document quotes, and all of them are stable.
What moved
The median: 1,875 ms → 3,553 ms.
The throughput: 116.2 → 51.3 tokens per second.
The error rate: 0.055% → 0.477%, which is still the best in the security line, and still a number from a gated card.
The volume: 24.59 billion tokens served over the earlier window, 21.53 billion over the later one [OURS]. Traffic did not collapse, which means the slowdown has not yet cost the card its users — or that they are batch workloads that do not notice.
The pattern is consistent with capacity: a card that onboarded a large pipeline between the two snapshots serves more tokens per second of wall clock and starts answers later. It is also consistent with a routing change nobody announced. Four days of data cannot distinguish them, and the honest description is “the shape moved, the reason is unknown”.
What the change costs in money
Nothing. Sequential throughput and median latency do not appear on an invoice — you pay per token, and the tokens cost the same.
What the change costs is everything else. A 30,000-token incident report at 116.2 tokens per second streamed for 258 seconds; at 51.3 it takes 585 seconds [OURS]. If your wrapper has a 300-second timeout — a completely ordinary number — the card went from comfortably inside it to failing on every long answer, with no change to the price and nothing in the UI to explain it.
That is the failure this kind of drift produces: not a bigger bill, a different error rate in your application. And it arrives without a changelog.
How to price an unstable shape
Three habits.
Re-measure weekly, and store the date with the number. One probe request per day is enough to see a factor-of-two move within a week. A latency figure without a date is an anecdote.
Set timeouts from your own p95, not the vendor’s p50. The gap between the two has been growing on this card.
Re-check before you commit to a deadline. A pilot that produced a throughput figure two weeks ago is describing a card that may no longer exist in that shape.
What to instrument so this is not a surprise
The drift was visible in the catalogue because the catalogue measures the card. Your application does not, unless you make it.
Log the first-token time per request. Not the total — the wait. A card that doubles its wait and keeps its throughput looks fine in an average and terrible to a user.
Record the model string from every response. If the name changes, the deployment changed, and that is worth knowing before the latency moves.
Alert on your p95, not the vendor’s. Your tail is the one that times out.
Re-run a fixed probe weekly. One request, same prompt, stored with a date. Four points is enough to see a factor-of-two move.
None of that needs a platform. A log line and a weekly glance at four numbers will catch the next 89% before your users do.
A note on the price that did not move
The tempting conclusion is that the card got worse for the same money. The accurate one is narrower: the card’s *serving shape* changed and its *price* did not, which is normal for a per-token product and abnormal for a per-request one.
A per-seat product that halves in speed is a refund conversation. A per-token product that halves in speed is a capacity event: you pay the same for the same tokens and wait longer for each one. The only place that shows up is in your client — timeouts, retries, and the queue behind them.
Which is why the monitoring habits above matter more than the complaint. Nobody is going to adjust the rate for a slower week, and the card’s benchmarks are unchanged. The only lever available to a buyer is the one that decides when a request is abandoned — and the only warning is a number you wrote down before it moved.
The number to carry away
3,553 ms — not because it is fast or slow, but because it is the card’s current shape and a timeout written against 1,875 no longer describes it. Re-measure before you commit, and date every figure you quote. A rate card is a contract and a latency profile is weather; only one of them is worth building a promise on.
The takeaway
OrcaCyber Zero 1.5 costs the same as it did four days ago and serves answers 89% slower with 56% less throughput, at a 0.477% error rate that is still the best in the line. Price it on $3 and $7.50, plan on 3,553 ms, set your timeout from your own tail, and re-measure — because the one thing this card has demonstrated is that its shape moves faster than its rate card.
Sourcing note: pricing, cache rate, context, output ceiling, latency, throughput, error rate and traffic figures are OrcaRouter production telemetry for the seven days ending 2026-11-05 [OURS]; the comparison uses the same gateway’s 2026-11-02 snapshot. Benchmark rows are OrcaCyber’s own published claims, unreproduced [VENDOR]. Access is restricted to authorized security research. Checked 2026-11-06.
