This cohort: 21 August – 26 August 2026
Get 5 days of unlimited frontier AI — with a chance to unlock 5× faster speeds.
How it works:
Last cohort, 16 August – 21 August: 9 people used Kimi K3 Abliterated free for the full window, and all 9 chose to roll themselves into this one.
This cohort adds Qwen 3.8 Abliterated with no token limits — the only limit is 12 concurrent sessions per seat. Kimi K3 Abliterated, which we run at 66 tokens a second, has proved to be as abliterated as the Qwen 3.8 models; see the comparison on our blog.
Refunds are processed on cohort boundaries only — 21 August, 26 August, 1 September, and each boundary after that. There is no refund part-way through a cohort, because a cohort is a month of hired hardware bought as a block.
We’re currently at 10/50 people — just 40 more spots needed to unlock 5× speed for everyone.
Same Kimi K3 abliterated, same 1M context, both sides of that arrow. You are not starting on a trial, a demo, or a smaller model — you get the full thing from the moment you join, at about 66 tokens a second, with unlimited tokens and the same answers. What the 50th seat buys is hardware: a dedicated 8x B300, and the same model replying at about 320 tokens a second instead. At 66 a long exploit chain is something you start and come back to; at 320 — 4.8x faster — it keeps up with you. Your month starts then too, not when you join, so nobody pays for time they cannot use at full speed.
You get NECROMICON today, and you are not charged unless 50 people join. Your card is authorized for the seat — reserved, not taken — and access starts immediately at 66 tokens per second. When the 50th paid seat lands we hire the box, the hold is captured, everyone steps up to 320 tok/s, and that is when your month starts counting.
Same price, same box, same seat pool. What differs is how much of yourself you hand over.
Access to standard-speed Kimi K3 abliterated and Qwen 3.8 abliterated (pingu-unchained-10) starts the moment you take a seat — whichever one you pick. Your card is authorized for $999, not charged, and NECROMICON is live for you straight away at about 66 tokens a second. You are only charged if 50 people join — and that is the same moment everyone here steps up to about 320.
$999 / month
Live immediately at ~66 tok/s
Sold by ALBUMERA LTD
Your card is authorized, not charged. Access starts straight away at ~66 tokens/sec — Kimi K3 Abliterated and Qwen 3.8 Abliterated (Unlimited). You are only charged if 50 seats are paid.
$999 / month
Live immediately at ~66 tok/s
Sold by ALBUMERA LTD
Your card is authorized, not charged. Access starts straight away at ~66 tokens/sec — Kimi K3 Abliterated and Qwen 3.8 Abliterated (Unlimited). You are only charged if 50 seats are paid.
$999 / month
Live immediately at ~66 tok/s
Sold by Audn Corporation
Your card is authorized, not charged. Access starts straight away at ~66 tokens/sec — Kimi K3 Abliterated and Qwen 3.8 Abliterated (Unlimited). You are only charged if 50 seats are paid.
There are no token limits and no overage charges on Qwen 3.8 abliterated based pingu-unchained-10 and KONG models. What there is, is a fair-share scheduler: nobody 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, never billed extra, and never silently moved to a smaller model. With 70 seats that ceiling is about 12% of the whole machine — which is why 70 is the ceiling and not a bigger number.
The money ones first, because those are the ones that matter before you decide.
No. We place a hold on your card for $999. A hold reserves the money against your available credit — it does not move it. Nothing is taken unless 50 paid seats are reached.
If the pool fills, the hold is captured and that is the moment you are charged. If it does not, the hold is cancelled and you pay nothing. How quickly the reserved amount reappears on your available balance is your bank’s decision, not ours — usually a few days.
The full model, immediately. Unlimited Kimi K3 abliterated at about 66 tokens per second — no token limits, no daily cap, the complete 1M context.
It is not a trial, a demo, or a smaller model. The only thing the pool filling changes is how fast it answers.
About 66 tokens/sec now, about 320 tokens/sec once 50 seats are paid and we hire a dedicated 8x B300. That is 4.8x on the raw numbers.
Being precise about it: 66 is a deliberate throttle, so it is always deliverable. 320 is what a seat reaches on a box that is not saturated — under heavy simultaneous load a fair-share scheduler applies, so treat it as up to 4.8x faster rather than a guaranteed floor.
When the box is hired, not when you join. The one-month clock starts the moment the 50th seat is paid and everyone steps up to the faster tier.
Waiting for the pool to fill costs you none of your month. That is deliberate — you should not pay for time you cannot use at full speed.
You choose, and the default costs you nothing:
We never charge you for saying nothing. Rolling over is something you opt into, not something that happens if you ignore an email.
Only on a cohort boundary — 21 August, 26 August, 1 September, and each boundary after that. Capture is the moment we commit to a month of 8x B300 rental, and that obligation does not shrink part-way through because one person changes their mind. The boundary is the exit.
The exceptions are all cases where we did not deliver: the pool misses its floor, or we fail to provide the box. Then you are refunded in full.
Same price, same model, same pool. What differs is how much of yourself you hand over:
The ID check is what buys the penclaw.ai features. If you would rather not do one, the API-only options cost the same and ask nothing of you.
Yes, and we will not pretend otherwise. Stablecoin payments settle immediately and cannot be reversed by us — there is no such thing as a hold on that rail.
So on the crypto option the money genuinely arrives. If the pool does not fill, we refund it in full to the address you paid from. That is why crypto seats are capped at 10: it is a refund obligation we can actually discharge.
We never train on your traffic, and prompts and completions are not retained.
Being exact rather than sweeping: request metadata — timestamps, token counts, which model, how much of the box you used — is retained, because the fair-use scheduler and billing are built on it. Content is not.
Card seats are sold by ALBUMERA LTD (United Kingdom). They appear on your statement as AUDN.AI.
Crypto seats are sold by Audn Corporation (United States), shown as AUDN CORPORATION.
It lowers the model’s latent tendency to refuse gray-area prompts during authorized security work. It is our own method, applied to Kimi K3.
The models on this cluster are effectively uncensored: a 0.2 to 2.5% compliance rate — how often the model still complies with its refusal training, where lower means more thoroughly abliterated. The method, the bench and the per-model results are on our blog.
It is meant for offensive-security work you are authorized to perform. It does not change what is legal, and it is not a licence to point it at systems that are not yours.
No. It is one month, one payment, and no subscription. When the month ends it ends — there is nothing to cancel and nothing that quietly charges you again.
To continue, you join the next pool.
$4 / $21per 1M input / per 1M output
You are billed for tokens, in arrears, and for nothing else. An agentic request runs about 50,000 tokens in and 200 out, which is $0.204, and 98% of that is the input side. Below roughly 47 million output tokens a month this is the cheapest of the three.
Cheapest right up until you start using it properly.
Start on platform.audn.ai$999authorized, not charged
Card seats are authorized, not charged: the money is reserved, not moved, and access is live immediately at about 66 tokens a second. Crypto seats settle immediately and cannot be held, so they are refunded in full if the cohort misses. If the 50th seat is paid we capture every card hold at once, hire the 8xB300, and your month starts then, not when you joined. If it is not, the authorization is cancelled by default and you pay nothing. If you don’t roll over your authorization to next cohort you lose access and the only day you can ask for refund is last day of the cohort. 26 August in this example.
Free while the cohort fills. Worst case 4.1x cheaper than metered above 47M output tokens a month.
$999per month, on either platform
Sold on penclaw.ai and on platform.audn.ai. These are two different products at the same price — each is $999 a month for unlimited tokens on the same cluster, and you buy the one whose surface you want. penclaw.ai carries the full engagement tooling and is the only route here that asks you for a government ID check; platform.audn.ai carries the OpenAI-compatible API and asks for no ID. Either is charged today and again on the same date each month, with no funding window to wait through.
For people whose next month looks like this one.
| Pay as you go | Cohort seat | Ultimate | |
|---|---|---|---|
| What you pay | $4 per 1M input, $21 per 1M output. About $0.204 per agentic request. | $999, once, for one month. | $999, every month. |
| When you are charged | In arrears, for tokens already spent. | Card seats: not until the 50th seat is paid, and then every hold is captured at once. Crypto seats settle immediately, before the cohort fills. | Today, and on the same date each month. |
| Commitment | None. Stop sending requests and it stops costing. | One month, and only if 50 seats are paid. | Month to month. Cancel and the next one does not run. |
| Speed today | About 66 tok/s on the shared tier. | About 66 tok/s from the moment you take the seat. | About 66 tok/s. |
| Speed once 50 seats are paid | No change. The cluster is hired by the cohort. | About 320 tok/s on the hired 8xB300, guaranteed to the 70-seat cap. | About 320 tok/s on the same cluster, guaranteed to the 70-seat cap. |
| Latency per request | 10 s to 5 min, depending on requests in flight across the cluster. | 10 s to 5 min. About 94 seconds worst case at the 50-user operating point. | 10 s to 5 min. About 94 seconds worst case at the 50-user operating point. |
| Concurrency | 12 generations in flight, in any mix of PenClaw and API. A 13th is rejected immediately rather than queued indefinitely. | 12 generations in flight, in any mix of PenClaw and API. A 13th is rejected immediately rather than queued indefinitely. | 12 generations in flight, in any mix of PenClaw and API. A 13th is rejected immediately rather than queued indefinitely. |
| When the queue is full | HTTP 429 with Retry-After. Never a silent drop, never billed. | HTTP 429 with Retry-After. Never a silent drop, never billed. | HTTP 429 with Retry-After. Never a silent drop, never billed. |
| Context window | 1,000,000 tokens. | 1,000,000 tokens. | 1,000,000 tokens. |
| Abliteration completeness | Effectively uncensored — a 0.2 to 2.5% compliance rate across these models. Research and results. | Effectively uncensored — a 0.2 to 2.5% compliance rate across these models. Research and results. | Effectively uncensored — a 0.2 to 2.5% compliance rate across these models. Research and results. |
| Usage while the cohort fills | Billed at the metered rate, as usual. | Free for Qwen 3.8 and Kimi K3 Abliterated until the cohort fills. | Included. You have already paid for the month. |
| If the cohort misses 50 | Nothing happens. You were never in it. | The authorization is cancelled by default and you pay nothing. Rolling over is opt-in, so ignoring the email costs nothing. | Nothing changes. The subscription runs on at about 66 tok/s. |
| Refunds | Nothing to refund. You pay for what you used. | On cohort boundaries only — 21 August, 26 August, 1 September, and each boundary after that. Never part-way through. Crypto seats settle immediately and are refunded in full if the cohort misses. | On cohort boundaries only — 21 August, 26 August, 1 September, and each boundary after that. None part-way through a month already started. |
| Renewal | Nothing to renew. | No. The month ends and it ends. There is nothing to cancel. | Yes. Every month until you cancel. |
| Best for | Under about 47 million output tokens a month. | Over about 47 million output tokens a month, for one month. | Over about 47 million output tokens a month, every month. |
| Availability | Now. | Now, standard speed until 50 paid seats hires the cluster. Selling closes at 70. Five-day window. | Now, inside the same 70-seat ceiling. |
| Who you contract with | ALBUMERA LTD, United Kingdom. Shown on your statement as AUDN.AI. | ALBUMERA LTD for card seats. Audn Corporation, United States, for crypto seats, capped at 10. | ALBUMERA LTD, United Kingdom if card. Audn Corporation, United States, if crypto. |
The flat seat is $999 whether the cluster is idle or saturated, so making it faster costs you nothing. Below is the same work, one user with three threads, eight hours a day, twenty-two days, priced at four different latencies. Here is what $999 buys you if you would be using pay as you go.
4.1x to 38x cheaper
4.1x is the WORST case, and it is the one to plan against: 94 seconds is the worst-case latency at the 50-user operating point with the cluster hired, and even there the same three threads over the same eight hours and twenty-two days cost $4,129 metered against $999 on a seat. Note which direction the metered column moves. At 10 seconds it is $38,814 — 38.9x the seat — because on a meter you pay for the speed you asked for.
While the cohort fills, that usage is free. Three threads at about 66 tokens a second is roughly 10.5 billion tokens a month, which would be about $42,700 metered. You are charged nothing for any of it unless 50 seats are paid.
And it moves as the cohort fills. Every seat sold after the 50th shares the same box, 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 — which is the same thing as saying the seat is worth less the fuller the cohort gets. Same one user, same three threads, same eight hours, same twenty-two days.
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.
| Users on the box | Latency, worst case | Latency, typical | Tokens per day | Tokens per month | Metered cost per 1M | Metered per month |
|---|---|---|---|---|---|---|
| 50 users, cluster hired | 94 s | ~24 s | 46.3 M | 1.02 B | $0.98 | $4,140 |
| 60 users | 112 s | ~28 s | 38.6 M | 848 M | $1.18 | $3,450 |
| 70 users, selling closed | 131 s | ~33 s | 33.0 M | 727 M | $1.37 | $2,957 |
So the seat is at worst 4.1x cheaper than metered at 50 users, about 3.5x at 60, and about 3.0x at the 70-seat close. It is never worse than the meter at this workload, and it never stops being $999.
4.1x is the lowest gain you will see. It is the worst case, at the worst seat count, on latencies our load tests put at four times the real figure. If you are going to move more than 47 million output tokens a month, the seat is a no-brainer — and everything above that floor only widens the gap.
| If you | Take | Why |
|---|---|---|
| You send under about 47 million output tokens a month | Pay as you go | 47 million output tokens is the breakeven against a $999 seat, at the$21 output rate. Below it the meter is cheaper, and you commit to nothing to find out. |
| You run agents all day and want the cluster for one month | Cohort seat | On a card the $999 is held, not taken; on crypto it settles immediately and is refunded in full if the cohort misses. Either way it is free while the cohort fills, and once the cluster is hired it is 4.1x cheaper than metered in the WORST case if you move more than 47 million output tokens a month — more than that at every lighter load. If 50 seats are never paid the authorization is cancelled by default and you pay nothing. |
| Next month looks like this month, and so does the one after | Ultimate | Same $999. The difference is that you get everything a cohort seat gets, for the whole month, guaranteed — with no funding window to wait through. It renews on its own, which is the point. A seat holder who takes the refund loses access to everything happening at Audn from that moment; on Ultimate the month is yours either way. |
Unlimited means no token cap, no daily quota and no overage charge on Qwen 3.8 abliterated based pingu-unchained-10 and KONG models. It does not mean no limits. There are four, and you should know them before you pay.
HTTP 429 with a Retry-After header. The request is not quietly dropped, not silently truncated, and not billed.And the number that matters most: about 320 tokens a second is the cluster’s physical output, and it is a guaranteed speed all the way to 70 users. That is precisely why we cap a cohort at 70 — the ceiling exists to protect the number, not to ration it. A cluster nobody else is on costs $47,000 a month, and it costs that whether one person uses it or fifty.
We hire the cluster from Modal and run our own abliterated weights on their DFLASH stack. The speed figures on this page 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:70 is where the curve stops being the number we promised, so 70 is where selling stops.
| Replica load | Speed per user | What that is |
|---|---|---|
| 4M tok/min | 668 tok/s | One stream, cluster otherwise quiet |
| 6M tok/min | 380 tok/s | Light concurrency |
| 8.7M tok/min | 235 tok/s | Approaching the seat cap |
| 10M tok/min | 210 tok/s | Replica saturated |
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 $47,000 a month, 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.