Responses API

from openai import OpenAI

client = OpenAI(
    base_url="https://api.vectoraxis.ai/v1",
    api_key="vk-c7d4e1f2a3b95a8b.Rq8mT2vX5nL9pW3kY7cJ1hF4dS6gB0zN2aE8uV5tQ1c",  # an Anthropic virtual key
)

response = client.responses.create(
    model="claude-sonnet-4-6",
    instructions="Answer in one sentence.",
    input="Why is the sky blue?",
)
print(response.output_text)

import OpenAI from 'openai';

const client = new OpenAI({
  baseURL: 'https://api.vectoraxis.ai/v1',
  apiKey: process.env.VECTORAXIS_KEY,   // vk-..., for any provider
});

const stream = await client.responses.create({
  model: 'gemini-2.5-flash',
  input: 'Write a haiku about gateways.',
  stream: true,
});
for await (const event of stream) {
  if (event.type === 'response.output_text.delta') process.stdout.write(event.delta);
}

Every VectorAxis feature is a request header, exactly as on chat completions: set them once with the SDK’s default_headers / defaultHeaders, or per call with extra_headers. See Request headers, and the API reference for the endpoint itself.

The decision is made per call, so a fallback or a routing config can move a Responses request from OpenAI to Claude and back: each provider gets it in the form it can take.

• Keys: BYOK and PLATFORM virtual keys, status, expiry, credit limit, model allowlist, and requests and tokens per minute and day. A raw provider key works on the direct-credential path: an OpenAI key unless x-provider names another provider.
• Plan limits: the workspace request rate.
• Guardrails: input and output, the key’s own plus any named in x-guardrail, streamed or not. A block is a 422 with x-guardrail-status: BLOCK and the error code guardrail_blocked.
• Tool-result scan: on every function_call_output, at the key’s own action.
• Gateway features: response caching (x-cache-mode; a hit streams as events too), retries and fallbacks across providers, routing configs (x-config), prompt attribution, logs and cost. Requests appear in Logs as Responses.

Errors come back in OpenAI’s shape, {"error":{"message","type","param","code"}}, so the SDK raises its usual exception classes; an OpenAI error keeps OpenAI’s own status and body. After the first stream event a failure arrives as an error event.

client.responses.retrieve, client.responses.delete and client.responses.input_items.list (GET / DELETE /v1/responses/{id}, GET /v1/responses/{id}/input_items) reach OpenAI, which stores the responses it creates unless the request sets store: false. A response from another provider is not stored anywhere, so it cannot be retrieved later.

Codex speaks the Responses API, so it routes through VectorAxis with a custom model provider in ~/.codex/config.toml. With an OpenAI key it runs OpenAI’s models as usual; with a key for another provider, set model to that provider’s model. See Coding agents.

model = "gpt-5.4-mini"
model_provider = "vectoraxis"

[model_providers.vectoraxis]
name = "VectorAxis"
base_url = "https://api.vectoraxis.ai/v1"
env_key = "VECTORAXIS_API_KEY"     # export VECTORAXIS_API_KEY=vk-...
wire_api = "responses"