Pricing
Priced on calibration, not on borrowed tokens
You bring and pay for your own model backend directly. CLI charges for calibration and monitoring, the part it actually does.
| Plan | For | Included |
|---|---|---|
| Developer | Evaluation and prototyping | Hosted API, unlimited calibration profiles of up to 5,000 examples each, community support. The offline statistics package is free and unlimited. |
| Team | Production workloads | Everything in Developer, plus drift monitors, label-efficient calibration, group-conditional profiles, and usage-based pricing. |
| Enterprise | Regulated, on-premises, or air-gapped | Statistics engine and calibration store in your VPC, zero-retention mode, backend-fingerprint audit trail, SSO, and dedicated support. |
Model-agnostic
Bring the model you already run
CLI reads whatever a backend exposes and picks the strongest valid nonconformity score for that access level. The guarantee shape stays the same; only its efficiency changes.
- OpenAIL0–L1
- Azure OpenAIL0–L1
- Anthropic ClaudeL0
- Google GeminiL0–L1
- AWS BedrockL0–L2
- OpenRouterL0–L1
- vLLM, self-hostedL1–L4
- SGLang, self-hostedL1–L4
- Your own modelL0–L4, declared
- L0
- Sampled text only
- L1
- Generated-token log-probabilities
- L2
- Scoring of text you supply
- L3
- Exact label-token probabilities
- L4
- Hidden states
With your own model, the SDK computes evidence locally and sends only probabilities, scores, and sampled answers. Read more about the access ladder and supported backends.
Calibrate the model you already have.
Start with the Python SDK against the hosted API, or run the statistics engine entirely offline.