Skip to main content
Frequently asked questions about models and their capabilities in Orq.ai. Each answer links to the page that covers the topic in full.
Browse the full catalog on Supported Models. Model IDs use the provider/ prefix, for example openai/gpt-5.6-sol, anthropic/claude-sonnet-5, or google/gemini-3.5-flash. Models must be enabled on the Models page before they can be called.
Requests route to the model provider, so provider credentials are required: connect BYOK keys for direct billing and rate-limit ownership, or use Orq.ai-managed access via Credits where available. See Billing & Usage.
Add models from private or self-hosted endpoints to the workspace model catalog. See Private Models.
The Models page lists the Context Length and Max Output Tokens columns for every model and can filter by context window size. The static Supported Models page does not list context windows or max output tokens.
Claude and most chat models require the conversation to include a user message. Requests whose input is only system or developer instructions, or empty, are rejected with a 400 invalid_input error. Tool results are grouped into user turns automatically.
Ranges are per model: Anthropic models accept 0 to 1, OpenAI models 0 to 2. Newer Anthropic models reject requests that set temperature and top_p together; set one or the other. Audio endpoints accept 0 to 1. See the Anthropic provider page and Reasoning.
Each provider exposes reasoning differently: OpenAI uses reasoning_effort, while Google Gemini and Anthropic use a thinking object. When reasoning_effort is set, the AI Gateway drops temperature and top_p before forwarding the request, since those parameters are incompatible with OpenAI reasoning models. See Reasoning.
Model providers reject incompatible parameter combinations, for example the Anthropic error 4028 when temperature and top_p are both specified. See Reasoning for the parameters the gateway drops on reasoning models.
Native file attachment works only with models that support file input directly. For Deployments, file input is supported on OpenAI, Anthropic, and Google Gemini models. PDF support varies by model. Use the routing rule “Keep file attachment requests on models that support them” to route by capability. See Multimodal and Files.
Image, PDF, and audio input support varies by model. Use the modality tabs on the Models page to scope models by type, or check the per-endpoint sections of Supported Models. See Multimodal.
Models that support structured outputs can return valid JSON matching a schema. Not all models support schemas; check the JSON Mode capability filter on the Models page and use JSON mode as a fallback. See Structured Outputs.
All models support streaming, with per-endpoint constraints. See Streaming.
Tool calling is supported by most chat models, with format differences per provider. See Tool Calling.
Prompt caching support varies by provider: Anthropic models use cache_control markers, Google Gemini models cache implicitly, and OpenAI models cache automatically. See Prompt Caching.
The Smart Router selects the optimal model from a pool per request based on task complexity and the chosen mode. For manual selection, compare models on the Models page.
Configure fallbacks and retries or load balancing so requests flow to another model automatically when the primary is unavailable.