How it works
The detail behind the club's promises: the two queues, why there's no daily limit, and how members choose the models.
For work you're waiting on
How priority works
- Your plan gives you priority tokens every day
5M, 10M or 20M, by plan. New prompt and output tokens count in full; cached prompt tokens count at a tenth, so a long conversation's history uses few priority tokens.
- Priority requests go first
They go ahead of all standard work, and switch the GPU back on if it's off overnight (that request waits a couple of minutes). When it's busy, the next free slot goes to whoever has the fewest priority requests running for their seats.
- Used them up? You carry on in the standard queue
Same models, no daily limit. You don't have to change anything: your requests move across on their own.
- It resets at midnight NZ
Tomorrow you're back on priority.
Priority tokens a day
Then the standard queue, with no daily limit, until midnight NZ.
For agents that run for hours
How the standard queue works
- No daily limit, and it never touches your priority tokens
Standard requests run whenever priority leaves room on a GPU. Each GPU works on 6 requests at once, and priority takes a free slot first.
- Send long jobs there yourself
Add
service_tier="flex"to a request, or use a key marked "Standard queue only". Use it for agents that can wait their turn, and keep your priority tokens for the work you're watching. - Give your tool a long timeout
When the club is busy, a standard request waits up to 10 minutes for a place. If none comes free, it's refused with a 429 and your tool should try again. Set a 15-minute timeout and plenty of retries:
OpenAI(timeout=900, max_retries=20). - When it's busy, members take turns
A member's turn lasts 5 minutes; then the next goes to whoever has used the least of the standard queue per seat over the last 30 minutes (Community has 1 seat, Pro 2, Pro Max 4). Each plan can also run a set number of requests at once (2, 4 or 5), so no one can take over while others wait.
- Nights
GPUs stay on from 7am to midnight NZ. Overnight, a GPU keeps going while there's work, and is switched off after 15 quiet minutes. Standard requests can't switch it back on, so they're refused until a priority request does, or the morning comes.
- Rare pauses
On a very busy night the standard queue may pause until 7am NZ, so the GPUs are ready for the day; priority requests still run. On a rare day the whole club may pause until midnight NZ. Either way your tool gets a 503 with a Retry-After that says when to try again.
One GPU, 6 requests at once
Filling the gaps
Why there's no daily limit
Priority work comes in bursts while people are at their keyboards, so a GPU often has room to spare. The standard queue fills that room with members' agents, which is why it needs no daily limit. Overnight, when nothing is running, a GPU is switched off.
One GPU's day (an example, NZ time)
Chosen by its members
Members choose the models
- Your seats are your votesCommunity has 1, Pro 2, Pro Max 4.
- A vote opens as the club growsAs seats fill, the club adds GPUs, and a vote opens on the next model.
- Members pick from new open modelsEach candidate shows what it's good at. Votes are weighted by seats. A bigger model may need more seats before it can run.
- The winner is added for everyoneAs soon as the club has the seats to run it.
qwen3.8-27bfrom openingIt's all in the open: the seats, the candidates, the votes as they come in, and every past result are on Models.
Reserve your seat for $0 until we open.
Priority tokens every day, and a standard queue with no daily limit.