What it costs

The five plans

Five plans, one principle: each of them stands for a model tier. Free costs nothing and works on the same tier as Starter; Enterprise is a contact plan and is arranged rather than selected.

Prices are per organisation and month, in US dollars. What a run consumes at the model provider comes on top — see the section after next.

What the plan decides

A plan here is not a feature list but a setting: the organisation's model tier. It is called Low, Medium or High, and it decides which model class does the work — and with it the quality and cost of every single run.

Tier and model class

In the hosted service, Low works with the Haiku class, Medium with the Sonnet class and High with the Opus class. The mapping sits in the deployment configuration: self-hosters set it themselves and can assign any model to any tier.

Ceiling and default at once

The plan’s tier is the organisation’s ceiling and its default at the same time: without another setting, every run works on it. A single ticket or a whole project may stay below it — never above, not even with a marker on the ticket.

Switchable at any time

The switch happens in the settings under billing and needs no conversation. If you notice mid-month that the tasks are harder than expected, you raise the tier.

Enterprise is arranged, not clicked

Enterprise is a contact plan: there is no tier to select, but a conversation about dedicated workers, the operating model and where the service runs.

A project can also leave the choice to the service (optional, off in the default delivery): it then starts on the cheapest rung and climbs only when verification fails — at most up to the plan's tier. That saves money on simple tasks without making the hard ones worse.

What comes on top

The plan price is not the model price. Every run consumes at the model provider, and what it consumed is recorded per run and shown in the billing view. A long, repeatedly repaired run costs more there than a short one — the plan's tier sets the price per call, not the number of calls.

Whoever stores their own model key settles directly with the provider. That does not depend on the plan: each organisation sets its key itself, and on the Claude paths the service requires it anyway — the security page describes how it is stored and what happens without it.

Cost ceilings for organisation, project and single run can be set. How they work and what applies without a set amount is on the operations page — in one place, so two pages do not tell two versions of the same thing.

Status of this page

This page is preliminary. It names the plans, their prices and the one thing a plan really decides today: the model tier. Deliberately, nothing more.

A paid plan today is arranged in a conversation: the organisation's owner sets the model tier in the settings, but there is no self-service checkout in the cockpit yet.

Volume quotas — runs, projects, sandbox time — are not part of a plan today. The service can enforce them, but in the default delivery it does not: without set values, unlimited applies. A number in a plan card would therefore be neither a claim for the customer nor a limit for us — just a number.

We also make no availability commitment. The service currently runs one instance of each data store, and such a commitment is a contract, not a marketing line. We write this down instead of omitting the line: whoever compares plans looks for exactly this gap, and finding it unnamed weighs more than reading it.

What changes, changes here first. If quotas or commitments arrive, they appear on this page before they appear on an invoice.