Models

Separate the model IDs accepted by configuration and shown in Agent Studio from the adapters that can actually run conversations.

Configuration and Agent Studio options#

The current agent configuration accepts these exact IDs, and Agent Studio lists all five in the LLM selector:

• gpt-5-mini
• gpt-5
• claude-haiku-4-5
• claude-sonnet-4-5
• gemini-2.5-flash

Being accepted by configuration or listed in the selector does not mean a production adapter is active. The models endpoint reports this distinction in adapter_active.

Production runtime availability#

Only gpt-5-mini and gpt-5 have a production adapter. They use the active GPT adapter when the Operade runtime has its OpenAI credential configured. Without that credential, a conversation also returns llm_unavailable.

claude-haiku-4-5, claude-sonnet-4-5, and gemini-2.5-flash are configuration and Agent Studio options, but their adapter_active value is false. Saving one of those IDs is allowed; trying to run a conversation with it returns llm_unavailable because no provider is wired to it.

Do not use selector presence as a runtime readiness check. Check adapter_active, then open a fresh test conversation.

Exact temperature behavior#

For gpt-5-mini and gpt-5, temperature is fixed at 1.0. Agent Studio disables the slider and resets the saved draft value to 1.0 when either model is selected. The configuration validator rejects another value with validation_failed. The temperature field is omitted completely from the provider request, including for a legacy configuration that contains another value, so the provider uses its own fixed default.

The Claude and Gemini options report temperature as adjustable, and configuration accepts values from 0.0 through 2.0. Those values are not sent to a provider today because those models have no active adapter; the conversation stops with llm_unavailable first.

Dashboard and API token controls#

The dashboard provides model and supported-temperature controls. max_tokens is an API field in the agent configuration; it is not a dedicated dashboard control. Its accepted value is -1 for unlimited, or an integer from 1 through 32,000.

This distinction matters when comparing a dashboard screenshot with configuration returned by the API: a valid field can be present without having its own visible control.

Change an active model safely#

1. Save the current expected answer for a small set of representative questions.
2. Select gpt-5-mini or gpt-5, confirm adapter_active is true, and wait for the draft to save.
3. Start a new test chat with the draft enabled.
4. Compare correctness, tone, and boundary behavior.
5. Publish only after the new draft passes those checks.

If the test returns llm_unavailable, confirm that the selected model has an active adapter and ask the Operade operator to check the runtime credential. Review availability and limits for the exact boundaries documented by this release.