Developer guide / A familiar API, an explicit contract
OpenAI-compatible,
with explicit boundaries.
Reuse supported SDK calls, not assumptions about the entire OpenAI platform. This is not a full drop-in replacement or a claim of equivalent model behavior. Evaluate your own workload before migrating.
1. Set all three connection values
- API key: create an UltimateModel platform key in API Access. Use
models:readfor model discovery and the capability manifest, andinference:invokefor inference. An OpenAI provider key is not a platform key. - Base URL: set
ULTIMATEMODEL_BASE_URLto your deployed application origin followed by/v1, not/api/v1. - Model: set
ULTIMATEMODEL_MODEL_IDto an ID returned byGET /v1/models. It identifies your serving-ready project, not an OpenAI model name or an arbitrary base-model ID. An empty list means no eligible model is visible to this key.
Store keys server-side in environment variables or your secret manager. Training completion alone does not establish serving eligibility: current artifact, license and evaluation evidence must pass the platform gates.
2. Try a supported request
Python · chat completions
import os
from openai import OpenAI
# Server-side only. Use your UltimateModel key, not an OpenAI key.
client = OpenAI(
api_key=os.environ["ULTIMATEMODEL_API_KEY"],
base_url=os.environ["ULTIMATEMODEL_BASE_URL"], # https://YOUR_APP_ORIGIN/v1
)
# Select a serving-ready model owned by your account.
models = client.models.list()
print([model.id for model in models.data])
response = client.chat.completions.create(
model=os.environ["ULTIMATEMODEL_MODEL_ID"],
messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)JavaScript · bounded text Responses
// Server-side Node.js only; never ship the key to a browser.
import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.ULTIMATEMODEL_API_KEY,
baseURL: process.env.ULTIMATEMODEL_BASE_URL, // https://YOUR_APP_ORIGIN/v1
});
const response = await client.responses.create({
model: process.env.ULTIMATEMODEL_MODEL_ID,
input: "Hello!",
});
console.log(response.output_text);Recorded offline official-SDK conformance: Python 2.48.0, JavaScript 7.4.0. These versions identify the evidence matrix; they do not establish compatibility with every SDK version or production availability.
3. Check the deployed capability manifest
The lists below are rendered directly from the canonical compatibility evidence shipped with this source revision. Before integration, request the live manifest using a platform key with models:read; check its version, expiry and state. This public guide does not bypass authentication for API calls.
curl "$ULTIMATEMODEL_BASE_URL/capabilities" \
-H "Authorization: Bearer $ULTIMATEMODEL_API_KEY"Manifest 2026-11-16.v3 · verified 2026-11-16 · expires 2027-02-16. An expired snapshot is not fresh compatibility proof.
Supported surface
models.listmodels.retrievechat.completions.non_streamchat.completions.ssechat.completions.stream_donechat.completions.usagechat.completions.system_messageschat.completions.tools_non_streamchat.completions.json_schema_non_streamresponses.non_stream_textresponses.message_arrayresponses.instructionsresponses.usage
Unsupported surface
embeddingsimagesaudiobatchesfilesassistantsresponses_streamresponses_toolsresponses_structuredresponses_stateresponses_backgroundtools_streamjson_schema_stream
Chat SSE is supported, but streamed tools and streamed structured output are not. Responses supports non-stream text and message-array input, instructions and usage—not streaming, tools, structured output, stored state or background work. Do not silently fall back to a different provider when these limits affect your application.
Handle failures before moving production traffic
Unsupported endpoints and request features fail explicitly with OpenAI-shaped errors rather than pretending to work. Unknown or inaccessible models return a model-not-found error. Authentication, key scopes, quota, artifact eligibility and provider availability can independently prevent a request. Check error status and code, keep retries bounded, and validate output quality and total cost with representative inputs.
Compatibility changes require a manifest version bump, a documented migration window, and passing conformance fixtures.