Models, providers and cost
This page is for administrators of your company. It explains which model answers questions on Orbit’s cloud, how many questions your plan allows, and, for installations run by a customer, how a model is chosen and how cost is kept at zero unless you decide otherwise.
On Orbit’s cloud, Lumen answers with Claude Haiku 5.5 from Anthropic. This is In development. You do not install or configure a model. How many questions an organization can ask is set by its plan; see Pricing. Answers still use only what the person asking may see, cite their sources, and change nothing without a confirm card. Find everywhere works without any model.
If the model is unreachable or a response is cut off, Lumen says so with a clear error.
Your own model on a customer-run installation Planned
Section titled “Your own model on a customer-run installation ”The model settings below describe installations run by a customer. Self-hosted Enterprise is Planned; see Self-hosted Enterprise.
On such an installation Lumen can use an open-weight model through an OpenAI-compatible endpoint. It costs 0 in tokens, and nothing is sent to a paid AI provider unless an administrator sets a provider key and a monthly budget above zero.
The paid provider (optional) In development
Section titled “The paid provider (optional) ”A paid provider such as Anthropic Claude can be added. The provider, the model and the effort level are chosen per function, with fallback models. Errors such as refusals, an exhausted token limit, 429 and 5xx are reported in clear words.
An administrator enters the key in the Lumen Admin tab. It is stored encrypted (AES-256-GCM) and shown masked. You can save and delete it. Use Check the connection to test the provider without generating text.
The budget: $0 by default
Section titled “The budget: $0 by default”The monthly budget for paid models is $0 by default. While it is $0, no paid request is sent. When a budget is set and used up, new paid requests stop and the user sees a clear message.
For every request Lumen records the tokens (input, output and cache) and the cost. An open-weight model’s price is 0.
The administrator console
Section titled “The administrator console”In the Lumen panel, administrators have an Admin tab with the items below. The provider, model, key and budget settings apply to customer-run installations Planned; on Orbit’s cloud the model is Claude Haiku 5.5, which is In development.
- the default provider;
- the Claude model and fallback models, plus provider, model and effort per function;
- the monthly budget and the amount spent;
- the Anthropic key (save, delete, masked) and Check the connection;
- conversation retention (0 to 365 days, 30 by default);
- switches to enable Lumen per product (a switched-off product is not offered to the model, not searched and gets no cards) and per project (inside a switched-off project Lumen does not answer, and its work items, repositories and pages are excluded from all reads and searches);
- a daily question limit per person;
- Quality checks: your own reference questions (see Administrator console and quality checks) and a Changes log of cards that were proposed, run or dismissed.
When nothing is configured
Section titled “When nothing is configured”If no model is available, Lumen does not pretend. The panel states what is missing, and on a customer-run installation what an administrator should do. Find everywhere keeps working, because it needs no model.
What reaches a provider
Section titled “What reaches a provider”Only the question, the page context, the rules and the manifest fields. See How Lumen answers and what it can see.