Caedral Models
Caedral-hosted products — current models, rates, and quota behavior.
Caedral Models are the products Caedral develops and hosts: Caedral Embed and native Caedral Voice. On a plan they are paid from your weekly credits first, then your prepaid balance. Hosted rerank runs on Caedral infrastructure but bills your prepaid balance. Requests use the same OpenAI-compatible key as the rest of the catalog.
Current models
| Product | Canonical ID | Owner | Pays with | Unit / rate | Availability |
|---|---|---|---|---|---|
| Caedral E1 Small (embeddings) | caedral-embed-e1-small-v1 | Caedral (hosted) | Credits → prepaid | $0.0025 per 1M tokens | Public — POST /v1/embeddings |
| Caedral Voice V1 | caedral-voice-1 | Caedral (hosted) | Credits → prepaid | $0.03 / 1k chars | Public — English, four voices, WAV 24 kHz |
The gateway accepts the canonical ID plus two aliases — caedral-embed and caedral/caedral-embed-e1-small-v1 — and every response echoes caedral-embed-e1-small-v1. Native Caedral Voice V1 (caedral-voice-1) bills $0.03 / 1k chars. Caedral-hosted reranking (BAAI/bge-reranker-v2-m3, $0.002 per request) runs on the same infrastructure and bills your prepaid balance. The branded caedral-voice product is third-party gpt-audio and also bills prepaid. See Specialized Models and /docs/caedral-voice.
Model card
Caedral E1 Small produces 384-dimensional L2-normalized embeddings over a 512-token context. Inference runs locally on an ONNX INT8 runtime. The model was developed by Caedral on an E5-small base with its own runtime and optimizations. Pass input_type as query, document, or passage for retrieval-aware prefixing; output dimensionality is fixed at 384 for the hosted model — other values return HTTP 400 unsupported_dimensions.
Wallet behavior
- Usage draws your weekly plan credits first and your prepaid balance second — it never blocks at the weekly credit boundary.
- Credits refresh every 7 days while the plan is active; the prepaid balance never expires.
- When neither wallet can cover a request, the API returns HTTP 402 insufficient_balance — top up or upgrade to continue.
- Every debit lands in the append-only billing ledger keyed by request id; GET /v1/usage summarizes the account.
OpenAI-compatible examples
curl https://api.caedral.com/v1/embeddings \ -H "Authorization: Bearer $CAEDRAL_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "caedral-embed", "input": ["Quarterly pipeline report", "Renewal risk notes"], "input_type": "document" }'import { Caedral } from "caedral"; const caedral = new Caedral({ apiKey: process.env.CAEDRAL_API_KEY }); const res = await caedral.embeddings.create({ model: "caedral-embed", input: ["Quarterly pipeline report"], input_type: "query",}); console.log(res.data[0]?.embedding.length); // 384SDKs
The official TypeScript, Python, Go, Java, and C SDKs all expose embeddings against this endpoint — point them at https://api.caedral.com/v1 and pass model caedral-embed (or the canonical ID). The raw REST route above works identically, and GET /v1/models lists the product alongside the external catalog so programmatic discovery needs no special casing.