Pricing
The code and weights are free. The hosted API is free for your first $20 of usage, then priced at what it costs us plus a flat margin. There are no output tokens to meter, because the model produces none.
Self-hosted
- Full model, training pipeline, evaluation suite and demos.
- Runs on a CUDA GPU, Apple silicon or CPU.
- Train on your own data with one command.
- No telemetry, no phone-home, no account.
Hosted
- Managed inference endpoint with the same
/v1/reviewcontract. - Managed fine-tuning on your JSONL: you keep the weights.
- Compute passed through at our provider cost, itemized on the invoice.
- Candidate evaluations are the only meter. Output tokens: none.
- Roughly a million 3-way decisions before the first invoice.
How the hosted meter works
Every review reports usage.candidate_token_evaluations: the total tokens the model actually scored across all candidates. That number, times the GPU cost per token on the hardware serving you, is the compute cost. We add 8% and bill monthly once you are past the free $20. There is no minimum and no per-seat charge. Rates change only when our provider rates change, and the invoice shows the pass-through line separately.
| What drives cost | Why |
|---|---|
| Number of candidates per question | Each candidate is one forward pass. A 3-way Choice is 3 passes; a 40-way is 40. |
| State length | The state is re-encoded for every candidate. Short states are cheap. |
| Questions per call | Cost adds across questions. Batching questions saves a round trip, not compute. |
| Output length | Zero. There is no generation. |
Why cost-plus
A decision model is infrastructure, and infrastructure should be priced like infrastructure: a known margin on a visible cost, not a per-seat number that hides the compute. The same rule applies across erp.ai products. If you can run it cheaper yourself, you should, and the self-hosted option is the complete product.
Questions
Is fine-tuning included? Hosted fine-tuning is billed the same way: GPU time at cost plus 8%. The resulting checkpoint is yours to download.
Do you train on my data? No. Your rows and your weights are used only to serve you. The public checkpoint is trained on public and synthetic data only.
What does $20 buy? On the order of a million short 3-way decisions on current GPU pricing. Exact counts depend on state length and candidate count; the usage field on every response tells you what you spent.
What happens at $20? Nothing stops. Add a card to keep going at cost plus 8%, or self-host the same model for free.
When does hosted open? Early access is by request while we size capacity. Request access above with your expected monthly decision volume.