Models & routing

You call the gateway with a logical model name, not a provider's raw model ID. The gateway resolves that name to a configured provider and upstream model. This keeps your client stable even when an administrator re-points a model behind the scenes.

Logical models

A logical model is a stable alias your administrator defines per project — for example:

Logical model	Typical use
coding-default	Day-to-day coding and reasoning
bulk	High-volume, lower-cost tasks

Send the logical name in the model field; the gateway routes it to the right provider and real model. If the administrator later swaps the underlying provider or model, your code does not change.

Which models can I use?

Your group determines the models you're allowed to call. You can see your allowed models on your Profile and on the Connect a client page. If you request a model your group isn't allowed, the gateway rejects the request.

TIP

Use the logical names your team agreed on (e.g. coding-default) rather than provider-specific IDs — it's the stable contract and lets the platform optimize cost and routing without breaking your client.

Automated model routing (auto)

If your administrator has enabled it, you can send the special model auto and let the gateway pick the right model from your prompt — a coding question routes to the coding model, a quick chat to a cheaper one — so you don't have to choose:

curl https://api.<your-domain>/v1/chat/completions \
  -H "Authorization: Bearer $KEY" \
  -d '{"model":"auto","messages":[{"role":"user","content":"summarize this article in two sentences"}]}'

If no route is a confident match, the request falls back to a default model. Everything else — budgets, limits, metering — works exactly as with a named model. Ask your admin which routes are configured for your project (they can show you on Projects → Auto Routing → Test a prompt).

Tool-result compression

If your administrator has enabled it for your project, the gateway may compact the content of tool messages before sending your request to the model — minifying verbose JSON and collapsing redundant whitespace. This shrinks the token bill on tool-heavy agent workloads.

It's transparent: you don't change anything in your requests. Only tool_result content is affected — your user, system, and assistant messages and the message structure are never altered. In the aggressive mode an unusually large tool output may be truncated to a head and tail with an honest …[N bytes elided by gateway]… marker in place of the middle. If you ever need the full, untouched tool output, ask your admin to set the project's compression mode to off or lossless.

Next steps

• Connect a client — send a request with a logical model.
• Administrators configure these in Routing and Auto routing.