Priced for the way you ship

Research is free once provisioned in private beta. Request access for the evals CLI; pay when Hyperpriors sits in the loop for production traffic.

01 — Tiers

Three ways in.

Research

For measuring before you ship. The evals CLI on your laptop, one project, and a community to compare notes with. Not self-serve yet — request access and we open a workspace.

Evals CLI · One project · Community

Free

Private beta · provisioned

Request access

Production

The full control plane for teams with users on the other end. Every provider, every instrument, and someone notified when a guardrail fires.

Full control plane · All providers · Slack escalation

Per seat

Billed monthly

Request access

Regulated

For teams whose auditors have questions. Immutable audit trails, data residency, and single sign-on across the plane.

Audit trails · Data residency · SSO

Talk to us

Annual agreement

Talk to us

02 — Comparison

What each tier holds.

Feature	Research	Production	Regulated
MeasureEvaluation before anything reaches a user.
Evals CLIGolden sets and suites, run anywhere	Included	Included	Included
Merge gatesFail the build when behaviour regresses	Included	Included	Included
Judge calibrationLLM judges checked against human labels	Not included	Included	Included
ProjectsSeparate systems under one account	1	Unlimited	Unlimited
Model providersOne interface across every provider we support	1	All	All
ControlThe runtime between your model and the world.
Harness runtimeTool mediation, bounded retries, escalation	Not included	Included	Included
Guardrail policiesBoundaries declared in code, enforced at runtime	Not included	Included	Included
Slack escalationUncertainty routed to a person, not a log file	Not included	Included	Included
Trace retentionEvery run replayable, byte for byte	7 days	90 days	Custom
AssureEvidence for the people who ask for evidence.
Audit trailsImmutable record of every decision and change	Not included	Not included	Included
Data residencyTraces stay in the region you name	Not included	Not included	Included
SSO / SAMLYour identity provider, everywhere	Not included	Not included	Included
SupportFrom forum thread to named engineer	Community	Priority	Dedicated

Every tier includes the same instruments where listed — no feature-gated dashboards, no artificial ceilings.

03 — Field note

“Clinicians didn't want a scribe that was confident all the time. They wanted one that knew when to say 'check this'. Once the drafts started flagging their own weak points, the conversation in our user interviews changed completely.”

A clinical documentation platformHealthcare technology

Read the case study

04 — Questions

Asked often, answered plainly.

01What does Hyperpriors actually do?

It is the control plane around your production AI systems: harnesses that mediate what a model can touch, evaluation suites that gate what ships, guardrails that enforce boundaries at runtime, and observability that records every step. Your models stay yours; we hold the layer that decides how much to trust them.

02Which model providers do you support?

Every major hosted provider and open-weight models behind one interface, including fine-tuned variants. Swap providers without rewriting your product, and diff behaviour across them in the same eval suite.

03Do prompts and traces leave our infrastructure?

Traces are stored encrypted, retained for your tier's window, and never used to train anything. On the Regulated tier they stay in the region you name, and self-hosted trace storage is part of the agreement.

04How long does installation take?

One command installs the CLI, and the first trace typically appears in under a minute. The harness wraps your existing calls; there is no rewrite, no proxy migration, and no SDK lock-in.

05Can we run evaluations in CI?

Yes — that is the intended home for them. Golden sets run as merge gates, so a pull request that regresses behaviour fails the build with a diff of exactly which cases moved and by how much.

06What happens when a guardrail fires?

Whatever the policy declares: block the output, flag it and continue, or route the run to a person. On the Production tier, escalations land in Slack with the full trace attached, so the reviewer sees what the model saw.

07Is this only for chatbots?

No. Most of our customers run pipelines, agents, and extraction workloads where no chat window exists. If a model's output reaches a user, a record, or another system, it can sit behind the plane.

08How is Hyperpriors priced?