Pricing
Pay for the loop, not for a pile of tokens.
Plans follow where the run lives and how many people touch the preference set. Usage is the labeled examples and the eval compute, not a blank pretraining corpus.
Lab
One researcher, one task
- A single active reward spec
- Preference budgets up to a research set
- Shared cluster runs
- Held-out eval gate
Team
A lab that labels together
- Yield by labeler and task
- Pair review before training
- Several concurrent runs
- Spec wiki shared across the group
Sovereign
Weights that cannot leave
- Lab machines and air-gapped hosts
- Bring-your-own checkpoints
- Named release reviewers
- Audit export for every shipped policy