The Uncensored Card That Costs Less Than the Model It Came From
On OrcaRouter the card sits beside a base that is more expensive per output token. Every abliterated card in the catalogue so far has been a premium product: the same weights, refusals removed, a higher price per token and a note about engagement and passkey. OrcaCyber Hy4, released 2026-11-04 onto OrcaRouter, breaks that pattern on both counts. It costs $0.80 in and $1.60 out per million tokens, while the preview it derives from — Tencent’s Hy4 preview — is $0.834 in and $2.501 out [OURS]. The uncensored version is 4% cheaper on input, 36% cheaper on output, and 20% cheaper on a 3:1 blend. It also carries a slightly larger output ceiling, 65,536 tokens against 64,000, on the same 1,048,576-token context.
That inversion is worth a piece on its own, because it breaks a rule buyers have learned to apply: “abliterated means pay more”. Here the premium is negative, and the reasons are visible in the two cards rather than in any announcement.
Why the price ratio can point the wrong way
An abliterated card is not the same product as a preview card, even when the weights are related. The preview is a vendor’s evaluation build — usually served conservatively, often on smaller capacity, and priced by the vendor’s own rate card. The OrcaCyber card is a re-served checkpoint with its own pricing, and there is no rule that says the buyer’s version must cost more than the vendor’s.
What the comparison actually shows is how the two were priced. The preview charges $2.501 for output and the abliterated card charges $1.60: that is not a discount on the same product, it is a different commercial arrangement for the same capability class. The input rates are nearly identical — $0.834 against $0.80 — which suggests the output rate, not the input rate, is where the preview’s margin sits.
For a buyer the arithmetic is straightforward. A workload of 100 million input and 20 million output tokens a month costs $83.40 + $50.02 = $133.42 on the preview and $80.00 + $32.00 = $112.00 on the abliterated card [OURS]. The saving is $21.42 a month, or 16%, on identical volumes.
What the same money buys in room and speed
The ceiling is the second surprise. The preview caps output at 64,000 tokens; the OrcaCyber card caps at 65,536 [OURS]. Neither number is large by the standards of the other security cards — K3.1 Pro offers 943,718 — but the abliterated card is not smaller than its base, which is the usual trade.
The speed gap is wider and easier to describe. Over the seven days ending November 5 the preview ran at a p50 of 4,232 ms and 31.8 tokens per second; the abliterated card ran at 3,613 ms and 114.2 tokens per second [OURS]. That is roughly 3.6 times the throughput at a 15% lower median latency, and it is the kind of difference that changes what a card can be used for: 31.8 tokens per second is a batch tool, 114.2 is usable in a loop.
Traffic says the same thing from the other direction. In three days the abliterated card served 8.00 billion tokens; the preview served 361 million in seven [OURS]. One of these is being adopted.

The number you must not read the usual way
The card’s seven-day error rate reads 19.462% [OURS]. That looks catastrophic and it is not the same statistic as a base card’s.
Every OrcaCyber card sits behind the `sec-uncensored` access gate: a request from an account without an engagement, a passkey and accepted terms is rejected before it reaches the model. Those rejections are errors, and they dominate the first days of a new card because most of the traffic is people testing whether they can get in.
The base card’s rate is a fair point of comparison from the other side. On the same day the preview reads 0.851%, and the base Kimi K3.1 Pro reads 15.506% while Claude Fable 5.1 reads 35.371% [OURS]. Rates in this range are describing capacity and access, not model quality. The honest way to read 19.462% is “gated traffic during three days of onboarding” — and the honest way to use it is to measure your own error rate after you are through the gate.

What is not on the card yet
No published benchmark rows. Three days in, the OrcaCyber Hy4 page carries price, context, ceiling, modalities and telemetry, and no score of any kind [VENDOR].
That absence matters more than usual here, because the reason to pay for an abliterated card is behaviour that benchmarks rarely measure. A user comparing this card with the preview has exactly one hard number — the price — and a set of properties: image input, tools, JSON, reasoning, a 65,536-token ceiling, and the gate. The rest is a pilot.
The pilot is cheap. At $0.80 and $1.60 a thousand requests averaging 3,000 input and 800 output tokens costs $2.40 + $1.28 = $3.68 [OURS]. That is the whole cost of finding out whether the card does the work, and it is less than the difference in a month’s bill between the two cards at the volumes above.
What a pilot would settle
The card has no published benchmarks, so the only way to answer the question it exists for — does it do the work — is to run your own.
Twenty tasks from last quarter, priced on the card: at 4,000 input and 1,000 output tokens each, $0.064 + $0.032 = $0.096 for the whole set [OURS]. That is a dime to find out whether a refusal-free card changes what your queue can produce.
Score three things per task: whether it completed, whether the output was usable without editing, and whether anything in it was wrong in a way a reviewer would have to unlearn. The third is the one that matters — a fluent but incorrect exploit description is worse than a refusal, because it sends the investigation somewhere.
Then compare against the preview, which costs 4% more on input and 56% more on output on the same tasks [OURS]. If the preview handles them, it is the cheaper and better-documented card; the abliterated one earns its place only on the tasks the preview declines.
The takeaway
OrcaCyber Hy4 is the first abliterated card in this catalogue that is cheaper than the model it derives from: $0.80 and $1.60 against $0.834 and $2.501, with a slightly larger ceiling and roughly 3.6 times the throughput. Its 19.462% error rate describes access-gate traffic in its first three days, not failures; the card still has no published benchmark, which is what a pilot is for.