Images and reranking

Per-unit pricing

Some models are not billed on tokens at all. Every catalogue row carries a price_unit: token means the per-token prices beside it are real, and anything else — image, second, document — means those are zero placeholders and unit_price_micros is what you will be charged.

Images

POST https://api.multigrid.ai/v1/images/generations, in OpenAI’s shape. The response adds a multigrid object naming the provider, how many images were billed and the per-image rate.

curl https://api.multigrid.ai/v1/images/generations \
  -H "Authorization: Bearer $MULTIGRID_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "black-forest-labs/flux-1-schnell",
    "prompt": "A cross-section diagram of a suspension bridge, technical illustration",
    "n": 1
  }'

Two image models are in the catalogue today — openai/gpt-image-1 and black-forest-labs/flux-1-schnell — and GET /models?modality=image is the list that cannot go stale.

Reranking

POST https://api.multigrid.ai/v1/rerank scores a list of documents against a query and is billed per document scored. It is the second half of a retrieval pipeline: /embeddings finds candidates cheaply, this orders them accurately. Indices and scores are passed straight through from the provider.

curl https://api.multigrid.ai/v1/rerank \
  -H "Authorization: Bearer $MULTIGRID_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "baai/bge-reranker-v2-m3",
    "query": "how do I cap spending on one key",
    "documents": ["…", "…", "…"],
    "top_n": 3
  }'

Video

There is a POST /video/generations endpoint on the same pipeline, billed per second of output. No video model is in the catalogue at the moment, so nothing can currently be routed to it — GET /models?modality=video returns an empty list, and that is the check worth making rather than taking this paragraph’s word for it.

Sending images in, not out

Attaching an image or a PDF to a chat request is a different thing and lives on /chat/completions. Use OpenAI’s image_url content parts. Limits are enforced before anything is sent upstream: 20 MB per image, 32 MB per PDF, 48 MB and 24 attachments across the whole request. A model that cannot read what you attached is dropped from the route list rather than being sent it — and if your fallback chain contains a visionVisionThe model can look at images and PDFs you attach, not just read text. model, that is precisely what the chain gets used for.