> ## Documentation Index
> Fetch the complete documentation index at: https://docs.platform.nora.my/llms.txt
> Use this file to discover all available pages before exploring further.

# Model defaults

> Workspace-wide model preferences — the default an Agent picks up when created.

Every new Agent block starts with a default model. Set the workspace's default here so you don't have to pick per-Agent.

## Setting the default

**Settings → Models → Defaults**:

* **Default text model** — used by new Agent blocks.
* **Default vision model** — used when an Agent's inputs include images.
* **Default voice model** — used for voice inputs.
* **Default embedding model** — used by Foundry Embed blocks and memory index spaces.

Existing Agents keep whatever model they were configured with — defaults only apply on creation.

## Per-Agent overrides

An Agent that needs a different model overrides the default via **Agent → Model** in the UI, or:

```bash theme={null}
nora agents update ag_1 --set-model gpt-4o
```

Overrides stick — the workspace default changing later doesn't affect Agents with explicit choices.

## Fallback models

Every Agent can also declare a **fallback model** — used when the primary hits its provider's limits or errors.

Configure per-Agent (there's no workspace-wide fallback). Common pattern:

* Primary: high-quality model (higher cost).
* Fallback: cheaper, faster model that's usually good enough.

The Trace shows which model actually ran per turn, so you can measure fallback rate.

## Cost defaults

Set the workspace default cost cap that new Agents inherit:

* **Default max cost per run** — dollar cap.
* **Default max tokens per run** — token cap.

Existing Agents keep their explicit values.

Prevents accidental "unbounded run" costs on new Agents that haven't been tuned yet.

## Choosing a default

Trade-offs:

* **Higher-tier default** (GPT-4o, Claude Opus, Gemini Pro) — new Agents work well out of the box. Higher cost until you tune down.
* **Lower-tier default** (GPT-4o-mini, Claude Haiku, Gemini Flash) — new Agents are cheap. You'll upgrade individual Agents when quality demands it.

Recommended: **lower-tier default, upgrade per-Agent based on simulation.** Cheaper baseline; explicit choice for anything premium. Nora's simulation view (see [Simulation](/reliable/simulation/overview)) makes the choice data-driven.

## When defaults matter most

* **Team onboarding** — new engineers building Agents don't have to decide upfront.
* **Bulk operations** — CI scripts that create many Agents don't need model choice per creation.
* **Templates** — Flow templates shipped to your org use the workspace defaults.

## Watching for drift

Some Agents get created and never tuned — they run on default forever. Health check:

```bash theme={null}
# Count Agents on the default model
default=$(nora settings models defaults --json | jq -r '.default_text_model')
for f in $(nora flows list --json | jq -r '.[].slug'); do
  count=$(nora flows get $f | jq --arg m "$default" '[.blocks[] | select(.kind=="agent" and .model==$m)] | length')
  echo "$f: $count Agents on default model"
done
```

High counts might mean untuned Agents worth reviewing.

## Related

* [Providers](/settings/models/providers) — the API keys behind the models.
* [`nora agents update --set-model`](/cli/agents) — override per-Agent.
* [Cost vs. quality](/reliable/simulation/pareto) — pick models by data, not guess.