GPT-5.6, from day one,25% cheaper.
All three GPT-5.6 models, plus Grok 4.3 and Kimi K2.7, behind one OpenAI-compatible URL. Point your SDK at it, and the signed receipt on every response proves nothing else changed.
Request · The entire migration
client = OpenAI(- base_url="https://api.openai.com/v1",+ base_url="https://api.draftworks.dev/v1",api_key=os.environ["OPENAI_API_KEY"],)
Response · Signed receipt
Drop-in compatible
Every major lab, one integration.
Current models from OpenAI, xAI, and Moonshot AI, served through the same endpoint and priced under their own list rates. The GPT-5.6 family was added the day it released.
- GPT-5.6 Sol
The GPT-5.6 flagship for reasoning, coding, and tool use. Succeeds GPT-5.5 at the same price.
1M contextTextNew - GPT-5.6 Terra
The mid-size GPT-5.6, close to flagship quality at half the price.
1M contextTextNew - GPT-5.6 Luna
The small, fast GPT-5.6 for high-volume and latency-bound work.
1M contextTextNew - GPT-5.4 Pro
Extended reasoning for demanding problems, served on the Responses API.
1M contextText - Grok 4.3
xAI's flagship, strong reasoning at a low unit cost.
200K contextText - Kimi K2.7 Code
Moonshot AI's open-weights model, built for agentic coding.
256K contextText - Kimi K2.7 Code Fast
The same K2.7 Code weights on faster serving, for interactive sessions.
256K contextText - GPT Image 2
OpenAI's image generation model, metered by token like every other model here.
ImageScale tier - GPT Realtime 2.1
Speech-to-speech sessions for voice agents, with text and image input.
VoiceScale tier
Cheaper, faster, provable.
The models you already run, served with less in front of them, priced under their own list rates, and verifiable down to the byte.
- Faster to first token
- The same weights the labs serve, with less sitting in front of them. Streaming usually starts in under half a second from us-east, and every response reports its own timing in
usage.latency_checkpoint. - draftworks~450msopenai direct~780ms
- Three labs, one account
- Models from OpenAI, xAI, and Moonshot AI share one API key, one credit balance, and one usage view. Moving a workload between them is a one-line change.
- Labs3Base URLs1Credit balances1
- Your prompts stay yours
- Nothing you send trains a model, and the Responses API
storeflag defaults to off. - Training on your dataNeverContent retentionNoneStored for billingHashes + counts
- A signed receipt on every response
- Some resellers quietly route traffic to a cheaper model. We could not if we wanted to: every response is signed with Ed25519 over the model id, token counts, and hashes of the exact bytes in and out.Verify a receipt
- {"model": "gpt-5.6-sol-2026-07-09","prompt_sha256": "9d2f…c41a","response_sha256": "41c8…0be7","total_tokens": 1590,"signature": "ed25519 ✓ valid"}
Pay for tokens. Nothing else.
Credits are prepaid inference: a dollar of credit buys a dollar of usage at your tier's rates, on any model. There is no subscription and nothing to cancel.
−20%
Base
Applies to every model from your first purchase, which starts at $5.
−25%
Scale
Reached automatically and permanently at $5,000 in lifetime purchases, and adds access to image generation and realtime voice.
gpt-5.6-sol $5.00 $4.00·gpt-5.6-terra $2.50 $2.00·gpt-5.6-luna $1.00 $0.80·input per 1M, Base tier
Full pricing and calculatorThe fair questions, answered.
Why can you be cheaper than OpenAI?
Draftworks serves the same models from committed capacity on Microsoft Azure, which costs less than retail API pricing. The spread funds the discount; nothing about the models or the API changes. The serving path is documented end to end in Architecture.
How are receipts verified end to end?
Every response is signed with Ed25519 over the model version, token counts, and hashes of the exact bytes in and out. Check any receipt in the browser on the verify page or in four lines of Node against the published public key, and audit the model itself by comparing logprob distributions with the vendor's own endpoint. How receipts prevent bait-and-switch walks through the whole argument.
What happens if upstream labs change terms?
Prices track each provider's published list, so your 20 to 25% discount holds whatever the list does. Credits are prepaid dollars rather than locked rates, capacity is committed with Azure ahead of demand, and model retirements are announced on the models page with API errors that name the successor. Data terms ride on Azure's data-processing agreements, covered in Data handling.
What do you store about my requests?
SHA-256 hashes of request and response bodies plus token counts, which is what receipts and billing need. Content is hashed and discarded, and the Responses API store flag defaults to off. The full lifecycle is in Data handling.