Unlimited Fast Cyber AI, weekly

Three ways to buy

All three, line by line

What a seat is worth as the cohort fills

And it moves as the cohort fills. The box is hired the moment the 10th paid seat lands, and every seat sold after that shares the same 8xB300 up to the 50-seat cap — so latency rises, the tokens one user gets through in a day fall, and the metered bill those tokens would have carried falls with them. The three rows below are the crowded end of the hired fast week; at the floor, with far fewer seats on the box, each seat gets comfortably more than these figures. Same one user, same three threads, same eight hours, across the seven-day fast week.

Every latency on this page is a worst case. Our load tests put the real figure at about a quarter of it — roughly 24 s where the table says 94 s. We publish the worst case because it is the number you can plan against, and because a seat that is already worth buying at the worst case does not need the average to make its argument.

Even at the crowded end, the fast week returns more usage than its share, and it costs nothing on top of your $1,000 of standard credits — whatever you run that week, you pay no meter for it.

The standard weeks run on your $1,000 of credits at the same $4.50 per-million meter as pay-as-you-go — no markup, no lock-in, and the balance is yours to spend across the month. The fast week is the part where flat beats metered; the standard weeks are simply metered use you have pre-paid.

Which one is for you

What “unlimited” means in the fast week

During the fast week, unlimited means no token cap, no daily quota and no overage charge on GLM 5.3 abliterated and Kimi K3 abliterated. It does not mean no limits. There are four, and you should know them before you pay. The standard weeks run against your $1,000 credit balance instead.

Twelve at once

Twelve generations in flight per subscription. PenClaw engagements and OpenAI-compatible API calls draw on one shared budget of twelve, in any mix. A thirteenth is rejected immediately rather than queued indefinitely, because a queue with no end is worse than a refusal. Twelve is a burst ceiling for spikes, not a reservation. The sustained figures above assume three.

10 seconds to 5 minutes

How long a request takes depends on how many requests are in flight across the whole box, not on what you are asking for. Near the 50-seat cap the crowded-box worst case is around 94 seconds; at the 10-seat floor, with far fewer seats sharing the box, it is a small fraction of that.

429, never a silent drop

Past the bounded queue you get an HTTP 429 with a Retry-After header. The request is not quietly dropped, not silently truncated, and not billed.

Ten seats’ worth

No single subscription may consume more than ten seats’ worth of the box. Above your share, while the box is busy, your requests queue behind requests below theirs. You are never refused for it and never billed extra.

And the number that matters most: up to about 666 tokens a second is the box’s streamed output on the fast tier, and the 50-seat cap is what protects it — the ceiling exists to hold the number up, not to ration it. A cluster nobody else is on costs about $11,750 a week — a quarter of the old $47,000 month — and it costs that whether one seat is on it or seventy.

Where the speed numbers come from

We hire the cluster from Modal and run our own abliterated weights on their DFLASH stack. The speed figures below are not ours: they are Modal’s published benchmark for moonshotai/Kimi-K3 on 8xB300, measured under the same agentic profile we actually serve — 50,000 tokens in, 200 out. You can check them yourself, which is the only reason a speed claim is worth anything.

Interactivity falls as the replica fills. That is the whole reason a cohort has a ceiling: 50 is where the curve stops being the number we promised, so 50 is where selling stops.

These figures apply once 10 seats are paid and the box is hired for the week. They do not describe what you get today. Until the floor is reached you are on the shared tier at about 50–70 tokens a second, on GLM 5.3 Abliterated and Kimi K3 Abliterated, spending from your $1,000 of credits. That band is a deliberate throttle rather than a point on the curve below, which is why it is always deliverable: it is a floor we hold, not a speed that degrades as more people join. The table starts applying the moment the 10th seat is paid, and not before.

Our abliteration is measured the same way, on the same weights we serve — method, bench and per-model results at blog.audn.ai. The cluster costs about $11,750 a week, which is why this is a cohort and not a button.

We found the sweet spots with a lot of research. Let’s have sovereign frontier AI together.

Specifications

Model

Kimi K3, audn abliteration

Context window

1,000,000 tokens

Modality

Text and native vision

Cluster

8x B300, us-east

Shared tier

About 50 to 70 tok/s. A deliberate throttle, so it is always deliverable.

Dedicated cluster

Up to about 666 tok/s (200–666 band). The box’s streamed output, not a per-seat allocation.

Concurrency

12 generations in flight per subscription, any mix of PenClaw and API

Latency

10 s to 5 min, depending on requests in flight across the box

Metered list price

$4.50 per 1M input, $4.50 per 1M output, $0.40 per 1M cached input