Skip to main content
Model Service provides one Aion-managed interface for calling supported models. It centralizes model discovery, authentication, permission checks, credit usage, and request attribution without requiring each application to manage a separate provider connection. Open Resources > Models to browse the current catalog. Each entry has an exact model ID, such as a provider-prefixed ID, that you pass to the API or an Aion SDK adapter.

​
Before you begin

You need:
  • an active Aion organization with Model Service access;
  • enough Aion credits for the request.

​
Authentication

Choose credentials based on where the caller runs:

​
Personal API key

Use a personal API key when calling the Model Service directly from a script, API client, or Codex. Open Settings > Personal > API Keys, create a key, and store it in your shell:
Send the key as a bearer token. It acts as your user in the active organization and is limited by your permissions. See API Keys for creation and security guidance.

​
Deployment client credentials

Aion deployments use AION_CLIENT_ID and AION_CLIENT_SECRET:
  • Remote Deployment: Aion displays both values after you create the deployment. Copy them immediately and add them to the runtime’s environment. The client secret is shown only once and cannot be retrieved later.
  • Aion GitHub: Aion injects both values into the hosted runtime environment automatically.
The Aion SDK reads these variables, exchanges them for a short-lived Aion token, and refreshes the token as needed. SDK model adapters use that token automatically; do not send the raw client ID or secret to a model endpoint.

​
Permissions and attribution

Both authentication paths require model.execute in the organization. A personal API key acts as its user. Within an agent request, the SDK preserves its request-local attribution, including a forwarded usage-attribution carrier. Outside a request, it sends no attribution header; Aion resolves the deployment’s current Daemon Identity. An active request with missing attribution fails instead of falling back to the daemon. Before sending work to a model, Aion also checks the organization’s credit policy. Usage records contain the model, token counts, Aion credits, result status, and responsible user or Identity. They do not contain the prompt or generated response. A failed or cancelled request can still use credits when the model reports consumed work.

​
Usage

Open Resources > Models and select a model. The Usage tab shows examples for its available operations. A model can have several output categories. A category describes its output, not the media that it accepts as input. Video, Audio, and Other entries can appear as metadata. A category alone does not enable an invocation endpoint. The model detail page explains when no operation is available. Chat model selectors continue to show text models only.

​
Choose a client

Use the Text client guide for SDKs, frameworks, and Codex. Use HTTP or the Python SDK example for Decisions. Use the image guide for JSON, SSE, and caller-owned image files.

​
Make an HTTP request

See the Text request example or the Decisions request example.

​
Use an SDK adapter

The Text SDK examples retain the existing adapter contracts.

​
Configure Codex

See the Codex configuration for the Text Responses endpoint.

​
Provider preferences and service tiers

Text requests support the existing provider preferences and service tiers. Decisions accepts a provider object but does not accept service_tier.

​
Provider preferences

See Text provider preferences.

​
Service tiers

See Text service tiers.

​
SDK examples

See the Text SDK examples with provider options.

​
Browse model details and pricing

The Models list shows each model’s provider, prompt and completion price per million tokens, recent token usage, and invocation count. Search the catalog or filter it by category and recent use. Models catalog showing search, reporting windows, model IDs, providers, pricing, tokens, and invocations Select a model to review its supported input and output modalities, context length, maximum completion size, knowledge cutoff, pricing, and other catalog metadata. The catalog is the source of truth for availability. Aion does not expose controls for enabling a model or configuring provider credentials on this page. Catalog prices are Aion customer prices in credits. A request uses the price policy for its operation and organization, even when the model has several categories. Prices without a verified operation remain unavailable. An unavailable price does not mean that a model is free.

​
Monitor usage and cost

Select a model and open Metrics to inspect Invocations, Prompt Tokens, Completion, Total Cost, and Principal Attribution. Use the 24h, 7d, and 30d controls to change the reporting window. Model metrics showing invocations, prompt tokens, completions, total cost, and principal attribution Model responses include usage.cost when Aion can calculate a quote in Aion credits for the returned result. Chat Completions, Responses, Decisions, and Images use this same convention. Fractional credits are valid. The quote uses the applicable organization policy when Aion returns the response. It does not include other billable attempts and is not a final settlement receipt. Policy changes or later usage information can change the settled charge. Metrics and billing records show settled charges in Aion credits. Decisions responses can omit token counts. Token totals include only available counts and can be incomplete. Invocation counts and settled Aion charges still apply. Missing cost information does not mean a request is free.

​
API behavior

Use stream: true or send Accept: text/event-stream to stream either Text creation endpoint. The Responses endpoint emits named events such as response.output_text.delta and response.completed. Chat Completions emits data frames and ends with [DONE]. The Responses endpoint is intentionally stateless. Set store to false, include the full conversation context in each request, and do not send stateful fields such as previous_response_id, conversation, or background: true.

​
Troubleshoot a request

For Decisions errors, see Decisions failure behavior.

​
What to do next