The questions people actually ask when a request comes back 429.
Trial keys share one concurrency pool. When the pool is full the request is refused immediately rather than queued, so you get 429 instead of a long wait. Retry in a few seconds. It is not an error in your client, and it is not a ban. Paid plans get priority in the pool.
The weights are open and we run them ourselves, so there is no frontier lab to pay for and no next release to amortise. That is also how we can offer 10× Claude Pro plan usage: we are not rationing a scarce frontier model, we are rationing one machine. What you give up is capability on the hardest problems, not room — see what runs here.
Set the model field, or ANTHROPIC_MODEL for clients that send one name everywhere. Valid ids: glm-5.3, deepseek-v4-pro, kimi-k3.1-pro. You can also pick a default in the dashboard; it applies to requests that arrive with a model name we do not recognise.
Yes. Set ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN. For OpenAI-shaped clients use the same key with base URL https://abliterateme.com/v1. Any model name is accepted — an unknown one just gets your dashboard default.
The oldest turns are dropped and the request still runs. You get a normal completion instead of an error, and usage.input_tokens reports what the model actually saw, so your context meter stays honest. This is what every provider does with a 1M session; we would rather say it than have you find it.
We keep aggregate usage metrics to run the service, and may retain transcripts briefly for debugging and abuse prevention. We do not train on your data.
Yes. There is no restriction on what the output is used for.
No. It is a small independent service, provided as-is. The status page is honest about incidents.
One key per account. The key is not secret from your own tooling, but anyone who has it spends your trial.