Skip to main content

​
Model format

Always provider/model-id: openai/gpt-4o, anthropic/claude-3-5-haiku-20241022, etc.

​
Parameter handling

The gateway reads per-model constraints from its built-in catalog (updated via pnpm dlx tsx packages/model-catalog/scripts/sync.ts). Before forwarding any request it:
  • Strips parameters the model doesn’t accept — temperature is removed for OpenAI reasoning models (o3, o3-mini, o4-mini) automatically
  • Clamps out-of-range values to the model’s min/max
  • Renames max_tokens → max_completion_tokens for reasoning models
You don’t need to know which models are reasoning models — just pass temperature and max_tokens as usual and the gateway handles it.

​
Deprecation warnings

The dashboard model picker shows deprecation notices for models with a deprecated_at date. Retired models are hidden by default.

​
Image and document inputs

Each model in the catalog declares whether it accepts images and documents (PDFs and other binary files):
  • Images — image_url and file blocks with image/* mime types
  • Documents — file blocks with application/pdf and other binary mimes
Most Google Gemini models, Anthropic Claude 3.5+ models, and OpenAI GPT-4o/GPT-5 models accept both — some variants (e.g. openai/gpt-5-nano, openai/o3) accept images but not documents. Vision models like mistral/pixtral-large-latest, groq/meta-llama/llama-4-maverick-17b-128e-instruct, and cohere/command-a-vision-07-2025 accept images only. Text-only models (DeepSeek, most Fireworks models) accept neither. Anthropic accepts images and PDF documents only; other binary document types are rejected. The gateway rejects requests whose content blocks the chosen model cannot process, and skips fallback candidates that lack the capability. The dashboard model picker shows the same warning while editing. Text files (text/*, CSV, JSON) are inlined as plain text and bypass these checks entirely. Model input files can be sent as base64 data URLs or remote URLs — see file inputs for details.