DCP exposes a drop-in /v1 API for chat completions and model discovery. The pages that show rates, model capability, context, and provider counts read the backend catalog, so zero-capacity models do not become marketing promises.
Loading live model metadata...
Use base_url=https://api.dcp.sa/v1 with your DCP renter key; SDK rewrites are not required.
Public model rows come from /v1/models and only count providers that are verified serving.
Model metadata carries SAR per-1M token prices, context, max output, and capability flags.
The shipped router policy is the balanced default; premium/latency/cost policies stay gated until measured.
Static-prefix and session hints are exposed as hash-only measurements; cached-input discounts stay off until settlement proof exists.
Loading prompt-cache readiness...
fig. 01 - verified model routingUse the same OpenAI SDK call shape, but keep your traffic on DCP's in-Kingdom provider mesh. The model catalog is the source of truth for what is actually serveable.
from openai import OpenAI
client = OpenAI(
api_key="$DCP_RENTER_KEY",
base_url="https://api.dcp.sa/v1",
)
response = client.chat.completions.create(
model="Qwen/Qwen2.5-14B-Instruct-AWQ",
messages=[{"role": "user", "content": "Explain zakat in Arabic."}],
)
print(response.choices[0].message.content)The Playground sends the balanced policy only when the backend marks it available. Cost-first, latency-first, premium, batch, and prompt-cache economics need measurement gates before they become public promises.
Loading router-policy readiness...
Serveable models, provider count, context, and token prices come from the live catalog.
OpenAI-compatible requests run through DCP's provider router and meter usage.
Prompt-cache rows are hash-only measurements; cached-input discounts and provider KV-cache control are still off.
Batch, prompt cache, LoRA, and dedicated deployment flags stay off until implementation and proof land.