Introduction
The GPT-5.6 family, Grok 4.3, and Kimi K2.7 through one OpenAI-compatible API. Served from Azure, priced 20-25% under list, with a signed receipt on every response.
Draftworks serves frontier models from OpenAI, xAI, and Moonshot AI through one OpenAI-compatible API backed by Azure's inference fleet. You point your existing SDK at a different base URL and keep everything else — request shapes, response shapes, streaming, tool calls.
https://api.draftworks.dev/v1Three things distinguish it from calling OpenAI directly:
- Price. Every model is 20% under provider list from your first dollar — GPT-5.6 Sol at $4.00 per 1M input tokens and $24.00 per 1M output. Once your lifetime purchases reach $5,000, rates drop permanently to 25% under list. See pricing and credits.
- Latency. Requests are served from Azure's inference fleet. Observed user-visible time-to-first-token is around 450 ms, and every response carries upstream latency telemetry in
usage.latency_checkpoint— surfaced by Azure and passed through unmodified. - Verifiability. Every response is signed with Ed25519. The receipt commits to the model id, request and response hashes, and token counts — so you can prove, cryptographically, that you got what you paid for. See verification.
Billing is prepaid credits: 1 credit = $1 of inference at your tier's rates. Minimum purchase $5. No subscription, no seat fees.
Who this is for
- Teams already built on the OpenAI SDK who want the same model at a lower unit cost, with zero code changes beyond
baseURL. - Anyone who needs an auditable record of inference — the model actually used, the exact bytes in and out — for compliance, research, or dispute resolution.
- Workloads where prompt content must not be retained. Draftworks stores SHA-256 hashes and token counts only; see privacy.
Endpoints
| Method | Path | Purpose |
|---|---|---|
| POST | /v1/chat/completions | Chat Completions — OpenAI-compatible |
| POST | /v1/responses | Responses API — OpenAI-compatible |
| POST | /v1/images/generations | Image generation — Scale tier |
| WSS | realtime.draftworks.dev/v1/realtime | Realtime voice sessions — Scale tier |
| GET | /v1/models | List available models |
| GET | /v1/receipts/{id} | Retrieve a signed receipt — public, no auth |
All authenticated endpoints take Authorization: Bearer dw_live_.... See authentication.
The models
Nine models from three labs, one catalog: the GPT-5.6 family (Sol, Terra, and Luna, added the day OpenAI released them), GPT-5.4 Pro, Grok 4.3, Kimi K2.7 Code in standard and fast mode, plus image generation and realtime voice on the Scale tier. Context windows, endpoints, and full rate tables are on the models page.
Embeddings, batch, and fine-tuning are not offered. Requests to those endpoints return 404 today. See SDK compatibility.
Next steps
- Quickstart — first request in under five minutes
- Models — the catalog and per-model rates
- Chat Completions reference
- Verification — how signed receipts work and how to check one