Migrating from OpenAI

The gateway implements the OpenAI wire format, so migrating an existing integration is a configuration change, not a rewrite. Three things change:

Base URL → https://api.smartapihub.com/v1
API key → your sk-llm-… key from the dashboard
Model id → prefixed with the vendor: gpt-4o-mini becomes openai/gpt-4o-mini

Before / after

typescript

import OpenAI from 'openai';

// before
// const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });
// after
const client = new OpenAI({
  baseURL: 'https://api.smartapihub.com/v1',
  apiKey: process.env.LLM_API_KEY,
});

const res = await client.chat.completions.create({
  model: 'openai/gpt-4o-mini', // was 'gpt-4o-mini'
  messages: [{ role: 'user', content: 'Hello' }],
});

python

from openai import OpenAI

# before: client = OpenAI()
client = OpenAI(base_url="https://api.smartapihub.com/v1", api_key=os.environ["LLM_API_KEY"])

res = client.chat.completions.create(
    model="openai/gpt-4o-mini",  # was "gpt-4o-mini"
    messages=[{"role": "user", "content": "Hello"}],
)

Most OpenAI SDKs honour OPENAI_BASE_URL and OPENAI_API_KEY, so you can migrate without touching code:

bash

export OPENAI_BASE_URL="https://api.smartapihub.com/v1"
export OPENAI_API_KEY="sk-llm-…"

Only the model ids in your code need the vendor/ prefix.

Model id mapping

OpenAI id	Gateway id
`gpt-4o`	`openai/gpt-4o`
`gpt-4o-mini`	`openai/gpt-4o-mini`
`text-embedding-3-small`	`openai/text-embedding-3-small`
`dall-e-3`	`openai/dall-e-3`
`tts-1`	`openai/tts-1`
`whisper-1`	`openai/whisper-1`

The exact set of available models depends on what your operator has enabled; always check GET /v1/models. Once migrated, switching to another vendor is one string: anthropic/claude-3-5-haiku, meta/llama-3.1-70b-instruct, and so on — the request shape stays the same.

What is the same

Request and response bodies for chat completions, completions, embeddings, images, speech and transcriptions.
SSE streaming format, including data: [DONE].
Tool calling (tools, tool_choice, tool_calls), response_format (json_object, json_schema), seed, penalties, logit_bias, stop.
Error envelope shape: { "error": { "type", "code", "message", "param" } }.

What is different

Area	Difference
`n`	Only `n: 1` is supported.
`usage`	Always present on non-streaming responses and always emitted as a final stream chunk — you no longer need `stream_options.include_usage`. Contains `x_llm_cost_micro` and, when estimated, `x_llm_usage_estimated: true`.
`provider`	Extra top-level field on responses naming the provider that served the request.
`routing`	Optional request extension to override routing per request. Stripped before forwarding.
Headers	`X-Request-Id`, `X-LLM-Provider`, `X-LLM-Model`, `X-LLM-Routing-Reason`, `X-LLM-Cost-Micro` on every inference response.
Errors	`error.code` uses the gateway's code enum, and `error.request_id` is always present.
Video / music	Async job endpoints (`202` + polling / webhooks) that do not exist in the OpenAI API. `Idempotency-Key` is required.
Responses API	Supported as a documented subset.
Fine-tuning, files, assistants, batches, realtime	Not provided by the gateway.

Checklist

Replace base URL and key.
Prefix model ids with the vendor.
Remove any code that assumes n > 1.
Read usage.x_llm_cost_micro instead of computing cost client-side.
Log X-Request-Id alongside your own request logs.
Handle 402 (credits/budget) in addition to 429 — see API Keys.