Point the OpenAI SDK at an OpenAI-compatible gateway in 3 steps

lili hami — Mon, 07 Sep 2026 03:22:17 +0000

Most teams don't need a new SDK for every model. They need a single OpenAI-compatible endpoint and a clean baseURL.

The 3 steps

Create an API key on your gateway.
Point the OpenAI SDK at the gateway base URL (usually ending in /v1).
Keep the rest of your code — chat completions, streaming, tools. Swap the model string when you want GPT, Claude, or Grok.

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_KEY",
    base_url="https://YOUR_GATEWAY/v1",
)

resp = client.chat.completions.create(
    model="gpt-4o-mini",  # or a Claude / Grok id your gateway exposes
    messages=[{"role": "user", "content": "ping"}],
)
print(resp.choices[0].message.content)

Footgun: Claude Code vs OpenAI clients

OpenAI SDK / most OpenAI-compatible clients want the base URL with /v1.
Claude Code / Anthropic-style clients often want the base URL without /v1 (they append /v1/messages themselves).

If Claude Code 404s on /v1/v1/messages, you double-prefixed the path — not an auth bug.

Longer walkthrough: OpenAI-compatible API guide and the short getting started guide.

What to actually test

Don't only hit chat completions. Check streaming, tool/function calls, long context, and the clients you really use (Claude Code, Codex, Cherry Studio). If those break, "compatible" is marketing.

Cost framing for agents

List $/1M tokens is a vanity metric. Retries, timeouts, and tool failures dominate. Measure cost per successful task.

If you want a managed OpenAI-compatible gateway for GPT / Claude / Grok: APICLAN — signup auto-credits $0.20 so you can make a real call, not a fake demo. USDT wallet top-ups; no card data stored.

DEV Community: lili hami

Point the OpenAI SDK at an OpenAI-compatible gateway in 3 steps

The 3 steps

Footgun: Claude Code vs OpenAI clients

What to actually test

Cost framing for agents