The LLM providers & models page lists every AI model available on your AGO instance. Models
registered here can then be selected as the default model for the platform or as a custom model on
an individual agent.
This page is reserved for superadmins. It appears under Settings in the admin menu only if your
account has superadmin rights.
Add a model
- Go to Settings > LLM providers & models.
- Click Add model.
- Choose a Provider from the list (OpenAI, Anthropic, Mistral, Google (Vertex AI), Azure
OpenAI, AWS Bedrock, OpenRouter, Groq, Cerebras).
- Choose a Model ID from the suggestions. If the model you need is not listed — for example a
model released after your AGO version — pick Other (enter manually) and type its ID.
- Adjust the Display name if needed. This is the name shown when picking a model for an agent,
and it must be unique.
- Click Save.
The provider is registered automatically the first time you add one of its models, so there is no
separate step to create it.
Model ID vs display name
- Model ID is the identifier sent to the provider’s API (for example
gpt-4.1). For Azure
OpenAI, this is the name of your deployment, not the underlying model.
- Display name is the label your team sees inside AGO (for example
GPT-4.1 (fast)).
Reasoning effort
Providers that support reasoning (OpenAI, Mistral, OpenRouter) show a Reasoning effort field.
Leave it on No reasoning unless the model you registered is a reasoning model — sending a
reasoning setting to a model that does not support it makes every call fail.
Azure credentials
Azure OpenAI models can carry their own Endpoint, API version and API key. All three are
optional: leave them empty and AGO falls back to your tenant’s Azure settings, then to the
instance-level Azure configuration. The API key is stored encrypted and is never displayed again —
the field stays empty when you reopen the model, and typing a new value replaces the stored one.
Edit a model
Click any row to open it. You can rename it, point it at a different model ID, change its reasoning
effort, or move it to another provider. Switching a model away from Azure removes its stored Azure
credentials.
Delete a model
Open the row menu and choose Delete. A model that is still in use cannot be deleted: the Used
by column shows where it is referenced (the platform default, fast or reasoning model, or a
specific agent). Point those settings at another model first, then delete it.
View token usage and estimated costs
Superadmins can open Token Usage to see daily usage and estimated USD costs by model,
with filters for date, model, agent and Usage type. The Conversation search and
Document indexing cards show their embedding input tokens and estimated costs separately.
Choose a usage type to narrow the daily table; the cards continue to summarize the selected
dates and model. Embeddings have no output-token charge.
Indexing costs cover document chunks and custom data rows sent to embedding providers, including
reindexing. Reusing an existing embedding does not add a charge. These costs are shared across
agents, so clear the agent filter to see them. They are not added to an individual conversation’s
cost and remain in the report after the source is deleted.
Search costs also include background knowledge retrieval. Embeddings used to deduplicate
questions appear under Question extraction.
The report also includes conversation tagging and evaluation, question extraction, memory,
knowledge processing and coverage, lab simulations and security evaluations, the AI debugger,
ticket processing, translation, background agents, and other AI tasks. Each category has its own
usage filter. Categories with recorded activity show a cost summary, including input and output
tokens and the number of calls. Summaries cover all usage types even when the daily table is filtered.
The Django admin LLM models > Token usage & cost report shows the same categories.
Conversation totals include linked tagging, evaluation, memory, question extraction, translation,
and background agent jobs, including jobs that run after the conversation ends.
Clear the agent filter to include shared jobs without an assigned agent.
Voice costs are excluded.
Costs use public list prices and exclude account discounts and free credits. Unavailable
means a model has no known price or a request has no recorded token count. Older embedding
requests cannot be priced retroactively. Indexing usage is recorded from the time this feature
is enabled; existing vectors do not provide historical token counts. Partial estimates include
only usage with a known price and token count.
Background AI costs are recorded from the time tracking is enabled. Earlier jobs cannot be
priced retroactively. A completed request still counts if its response cannot be processed;
rerunning a job can therefore add more calls and cost. 