Generation controls
Use only the controls enabled for your selected model.
Updated October 31, 2026
Capability comes first
The request schema describes accepted shapes. The published model capability determines whether a control is currently available. Unsupported controls are rejected before a reservation or inference dispatch; there is no silent fallback.
Effort
Chat Completions and Responses accept effort values off, low, high and max. The reasoning_effort alias also accepts none as off; reasoning.effort is another supported shape. Conflicting aliases are rejected. An accepted effort value does not establish live support for every model or alter its output ceiling.
Sampling and formats
Where enabled, temperature is 0–2, top_p is 0–1, and stop accepts at most 16 strings. Response format controls include text, JSON object and JSON schema; Responses also accepts text.format. Read the generated endpoint schema for the exact object shape. A JSON format request does not replace application-side validation.
Visible reasoning
Reasoning text or a summary is returned only when that model capability has been explicitly enabled and the service actually receives it. NotHoly does not fabricate hidden reasoning. Messages keeps its separately documented supported subset; Chat and Responses controls do not imply equivalent Messages behavior.