API reference

Models catalog

Every serverless model, pricing, and context window. Pull it live from the API.

Live catalog endpoint

curl https://api.runaii.cloud/v1/models

{
  "object": "list",
  "data": [
    {
      "id": "glm-5.3",
      "object": "model",
      "context_length": 202752,
      "pricing": { "prompt": 1.40, "completion": 4.40, "input_cache_read": 0.26 }
    }
  ]
}

Pricing is USD per 1M tokens; input_cache_read appears on models with prompt caching. Integrate against this endpoint rather than hardcoding rates — the catalog is the source of truth and prices can change.

Currently served (catalog honesty pass, 2026-10-19)

Model	Context	Prompt $/1M	Cached $/1M	Completion $/1M	Best for
`glm-5.3`	202,752	$1.40	$0.26	$4.40	Frontier agentic + coding, tool calling
`deepseek-v4-flash`	1,048,576	$0.22	$0.022	$0.66	1M-ctx workhorse — long docs, cached prefixes
`qwen3.8-flash`	32,768	$0.40	$0.04	$1.60	Fast multilingual lane

Context values are the deployed serving limits, not model-name maxima.

Catalog policy (2026-10-06)

The catalog lists ONLY models verified serving at publish time — a model whose fleet is parked or whose lane has no route is removed, not listed as "unavailable". When a lane comes back (e.g. the dsv41 fleet un-parking), its model re-enters the catalog with pricing in the same regeneration pass (config/models.manifest.json → node scripts/generate-registry.mjs). If you need a model that is not listed, it cannot be called today: requests to non-catalog ids return 404 model_not_found.

A model is only callable when it appears in GET /api/v1/models; each entry also carries a live is_ready flag — a false flag means the lane is warming or not yet integrated, and calls against it will not serve. Check the flag (or Console → Models) before wiring a model into a production path.

OpenAI / Anthropic compatibility

OpenAI SDK: swap base_url to https://api.runaii.cloud/v1 — done
Anthropic SDK clients: point base_url at https://api.runaii.cloud/api — POST /api/messages accepts the Anthropic /v1/messages dialect and translates automatically (tools, tool_choice, tool_result round-trips, image blocks, top_k, and thinking → reasoning included). Tool-use and image blocks require an API key — the browser-console session path is text-only.
Model ids, pricing, max_tokens caps, and rate limits are identical on both dialects — there is no "cheaper" or "faster" path, only the SDK you prefer