Change your default model on Customer Key

On Customer Key (BYOK), you choose which models Glean uses by default. To change a default model:

1. Confirm you have enough provider capacity
2. Add the model to your LLM configuration
3. Set the model as your default
4. Monitor after the change

This page uses the GPT-5.6 Luna model as the example. The same steps apply to any model you make a default.

note

This page applies to Customer Key (BYOK) deployments. On the Glean Universal Model Key, Glean selects default models for you and there's no org-wide default for admins to set. See Model choice.

Prerequisites

Confirm each of the following. Each one prevents a common failure.

• Your hosting provider serves the model. GPT-5.6 Luna is available in Glean through OpenAI and Azure OpenAI. Check Supported LLMs before you plan a migration.
• Your text defaults sit with one provider. The large, small, agentic, and fast agentic defaults must all come from the same hosting provider. Image generation and realtime voice models are the exception and can use a different one. See Feature availability by LLM key, provider, and cloud.
• You can open the LLM insights dashboard. You need the Insights Moderator role. The capacity check depends on the usage numbers it reports.

Confirm you have enough provider capacity

A token-per-minute limit set below your real peak is the most common cause of HTTP 429 errors after a model change, and the errors start as soon as production traffic reaches the new model. To confirm your provider capacity, follow these steps:

1. In the Insights dashboard, open LLM.
2. Read Peak TPM and Peak RPM across a window wide enough to include your real spikes, not just the past week. Rate limits bind at the peak, so the average will mislead you.
3. Compare those numbers against the tokens-per-minute and requests-per-minute limits configured in your OpenAI or Azure OpenAI account for the model you plan to make a default.
4. Check the limits on every other model still carrying traffic. They draw on the same account-level quota.
5. If your configured limits don't sit above your observed peak with headroom, raise them in your provider account before you continue.

Glean publishes baseline capacity requirements by user count for OpenAI and Azure OpenAI. Treat those as a floor and size above your own observed usage.

note

TPM and RPM values in the dashboard are estimates from Glean's internal instrumentation, intended for capacity planning rather than billing reconciliation. Quota also isn't a guarantee of throughput. If your workload is latency-sensitive or spiky, ask your provider about provisioned capacity.

Add the model to your LLM configuration

You must configure a model before it appears anywhere in Glean, including as a default.

1

Open your LLM configuration

Go to Admin console → Models.

2

Add the provider

Click Add LLM, then select OpenAI or Azure OpenAI.

3

Enter your credentials

For OpenAI, enter the API key from the OpenAI platform. For Azure OpenAI, enter the key and endpoint from Keys and Endpoints in the Azure portal.

4

Select the model

Select GPT-5.6 Luna.

On Azure OpenAI, also enter the deployment name you configured for the model. Glean needs the deployment name, not the model name, and it has to match your Azure resource exactly.

5

Validate and save

Click Validate, then Save.

For the full provider configuration reference, see OpenAI setup or Azure OpenAI setup.

Set the model as your default

After you confirm your capacity and add the model, select it as one of your default models.

1. Go to Admin console → Models.
2. Select the provider you added when you configured the model.
3. Select GPT-5.6 Luna for each default you want it to serve, such as the agentic model or the fast agentic model.
4. Click Validate, then Save.

Move your other text defaults to the same provider at the same time.

Monitor after the change

In the Insights dashboard, watch three signals over the first day or two:

• 429 rate. A sustained increase means your provider limit is too low. Raise the limit in your provider account.
• 5xx rate. This separates provider-side instability from quota pressure. Check your provider's status page when it climbs.
• Response quality. Spot-check the query types your teams run most often.

The Usage breakdown table attributes 429 and 5xx rates per model, which tells you whether the new default is the model under pressure or something else is.

important

The LLM insights dashboard updates once daily and can lag by up to 24 hours. It won't confirm a healthy switch in real time, so don't read an empty chart as a clean result. For faster signal, watch user-reported errors and your provider's own metrics.

Roll back

To revert the default model setting, go to Admin console → Models, select your previous model for each default you changed, then validate and save. Keep the earlier model configured until you're confident in the new default, so a rollback doesn't require re-entering credentials.

Troubleshooting

What you see	Likely cause	What to do
The model isn't in the Add LLM model list	Your configured provider doesn't serve it. GPT-5.6 Luna is offered through OpenAI and Azure OpenAI only.	Add OpenAI or Azure OpenAI. On Customer Key you can't reach a provider outside your own cloud.
The model isn't available as a default	The model was never added to your LLM configuration, or validation failed	Confirm the model is saved under OpenAI or Azure OpenAI and that validation succeeded.
Validation fails on Azure OpenAI	Deployment name mismatch	Enter the deployment name from your Azure OpenAI resource, not the model name. See Troubleshoot LLM provider errors.
HTTP 429 errors after the switch	Your provider limit is below peak demand	Pull Peak TPM and Peak RPM from LLM insights, then raise the limit in the Azure portal under Quotas, or in the OpenAI dashboard under Settings → Organization → Limits.
Higher cost than expected	Text defaults split across providers, or older models still carrying traffic	Confirm your large, small, agentic, and fast agentic defaults share one provider, then review per-model usage in the Usage breakdown table.

If you encounter an issue that's not covered here, contact your Glean account team with your deployment name and the timestamps of the behavior.

See also

• Supported LLMs
• Configure LLMs in the Model Hub
• LLM insights
• Model deprecation
• Troubleshoot LLM provider errors