The AI Model Catalog

Every AI model available to your agents and the Luna assistant is listed in one catalog, and every model is labelled with where it processes data, which company operates it, and how capable and expensive it is.

Overview

Agents and the Luna assistant in Orbit MissionControl run on AI models from the model catalog. The catalog is managed under Settings → AI Models and has two tiers:

Standard — available to every workspace. Processing takes place in the EU.
Extended — additional models from other providers. An administrator activates the extended catalog for the workspace; depending on the model, processing can take place outside the EU, and the operating company can be based in the USA or in China.

Key highlights

Each agent can use its own model; the Luna assistant has a workspace-wide chat model.
Every model carries three transparency tags: processing location, operating company, and data retention.
Two indicators — intelligence and price — rate every model from 1 to 5 on one shared scale, so any two models are directly comparable.
Switching the extended catalog off moves affected agents back to the standard model automatically; nothing breaks.

The Tags on Each Model

Every model in the picker shows three separate facts. They are deliberately separate, because they answer different questions.

Processing location

EU Processing — requests are processed in the EU. Possible regions include Frankfurt, Ireland, Paris, Stockholm, Milan, and Spain.
US Processing — requests are processed on infrastructure in the USA.
Region Not Guaranteed — no processing region is contractually guaranteed; processing can take place outside the EU.

Operating company

The company that operates the model's inference — for example US Company or Chinese Company. This is about the legal entity, not the servers: a company based in China can operate a model from data centres in the USA, and a company based in the USA can process in the EU. That is why the location and the company are shown as two separate tags. The tag always names the operating company's jurisdiction, whatever the physical location says.

Data retention

No Data Retention — request data is not stored and is not used for training.
Retention Possible — request data can be stored for a limited time, depending on the operating company. It is not used for training.

The Intelligence and Price Indicators

Beside every model, two meters rate it from 1 to 5:

Intelligence reflects the model's capability, assessed on a public benchmark comparison (the Artificial Analysis Intelligence Index). The value is a periodically refreshed snapshot, not a live figure. Higher means stronger.
Price reflects the billed blended token rate, including any surcharge for extended catalog models. It is one scale across both tiers: a standard model rated 4 really is more expensive per token than an extended model rated 1. Higher means more expensive.

The price indicator compares token rates, not invoices: a cheap model used heavily can still cost more per month than an expensive model used rarely.

Data Processing and Subprocessors

Activating the extended catalog is a workspace decision taken by an administrator. Depending on the models your team then selects, request data is processed by additional subprocessors. The current subprocessor list, including each operating company's legal entity, purpose, location, and transfer safeguard, is published at orbit.do/subprocessors. No model in the catalog uses your request data to train AI models.

Example

Julia, an operator in Lyon, sets up an agent that summarises incoming bookings. Volume is high and the task is simple, so she picks a model rated 1 on price and accepts a modest intelligence rating. For the Luna assistant, her administrator Mohammad keeps a model tagged EU Processing with a high intelligence rating: conversations are low-volume but quality-sensitive. When their compliance team asks where data goes, the tags answer it directly — and the subprocessor list carries the details.

FAQ

Can each agent use a different model?
Yes. The model is set per agent; the Luna assistant has its own workspace-wide setting.

What happens to agents on an extended model if we switch the catalog off?
They keep running — on the standard model instead. The original selection stays visible in the picker, marked as unavailable, and takes effect again if the catalog is reactivated.

Is our data used for training?
No. No model in the catalog uses request data for training. The retention tag tells you whether data can be stored at all.

Why does a model show EU processing but a US company?
Because both facts are true at once: the servers are in the EU, and the company operating them is based in the USA. Showing only one of the two would hide the other.

Does a higher price indicator mean a higher invoice?
Not necessarily. The indicator compares per-token rates. Your invoice depends on usage: how often agents run and how much text they process.