From OpenAI

The three changes

1. Base URL https://api.openai.com/v1 becomes https://api.ai.ml/v1.
2. Key sk-… becomes aiml-live-… (create one in the console, inference scope).
3. Model gpt-5-mini becomes openai/gpt-5-mini. Every model is vendor/model; a bare OpenAI model name is also accepted and resolves to openai/<name>.

const client = new OpenAI({
  baseURL: (process.env.AIML_BASE_URL ?? "https://api.ai.ml") + "/v1",
  apiKey: process.env.AIML_API_KEY,
});

What stays the same

• Chat Completions request and response bodies, including tools, tool_choice, response_format (json_object and json_schema), stream, stream_options, n, seed, logprobs, stop, penalties and user.
• Streaming frames (chat.completion.chunk, data: [DONE]).
• The error envelope {"error":{"message","type","param","code"}}. code carries the aiml code, which is more specific than OpenAI's; see Errors.

What is different

Steering the route

The optional aiml object on the request carries what the OpenAI format cannot express:

{
  "model": "openai/gpt-5-mini",
  "messages": [{ "role": "user", "content": "hi" }],
  "aiml": {
    "route": {
      "providers": { "order": ["openai", "azure"] },
      "require": { "regions": ["eu"], "zdr": true, "max_price": { "output": 2.5 } },
      "sort": "latency",
      "fallbacks": true,
      "ttft_budget_ms": 1500
    },
    "metadata": { "app": "support-bot", "tags": { "tier": "free" } }
  }
}

route.providers (only, ignore, order) and route.require (regions, zdr, data_policy, max_price in USD per million tokens, reasoning) narrow the candidate endpoints; sort, fallbacks, affinity, stream_retry, schema_retry, ttft_budget_ms and max_duration_ms shape how they are tried. metadata (app, user_id, tags) lands in analytics and on the generation record.

The Responses API

/v1/responses takes the same body the OpenAI SDK sends (client.responses.create): input as a string or items, instructions, tools (function and hosted: web_search_preview, file_search, code_interpreter, image_generation), tool_choice, text.format, reasoning, include, max_output_tokens, temperature, top_p, stream, user, metadata. Streaming returns the named events (response.created … response.completed) with sequence_number; the final response.completed object carries usage and aiml.

curl -sS "$BASE/v1/responses" \
  -H "Authorization: Bearer $AIML_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"model\":\"${AIML_MODEL:-openai/gpt-5}\",\"input\":\"Say hello in one short sentence.\",\"max_output_tokens\":64,\"store\":false}" \
  | python3 -c '

The gateway is stateless, which changes three things:

Every model behind the gateway is reachable through /v1/responses, including providers that speak another wire format: items translate to the IR and back, so input_image, input_file, function_call / function_call_output and text.format: json_schema work against Anthropic and Gemini endpoints too, with aiml.warnings listing anything the endpoint could not honour.