Moving an existing OpenAI integration to one endpoint

By the aitokensai editors · Verified 2026-10-27 Sources: OpenAI API docs

The good news about migrating off the OpenAI API is that the OpenAI wire format won. Most of the ecosystem speaks it, so a gateway that accepts it means your client code largely does not change You change where it points and what it is allowed to ask for.

Three fields, in practice: the base URL, the key, and the model IDs.

Before: what an OpenAI call looks like

from openai import OpenAI

client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

r = client.chat.completions.create(
    model="gpt-5",
    messages=[{"role": "user", "content": "Summarise this"}],
)

After: the same call through one endpoint

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["GATEWAY_KEY"],          # one credential
    base_url="https://aicomp.ai/v1",            # one destination
)

r = client.chat.completions.create(
    model="gpt-5",                              # same model ID
    messages=[{"role": "user", "content": "Summarise this"}],
)

That is the whole diff for the common case. The SDK, the method names, the message shape, and the response object are unchanged, because the endpoint speaks the same protocol.

// Same change in Node
import OpenAI from "openai";
const client = new OpenAI({
  apiKey: process.env.GATEWAY_KEY,
  baseURL: "https://aicomp.ai/v1",
});

What actually changes

Model IDs. Your endpoint serves what it serves, which is not the same list as one vendor's. Pull it rather than assume it:

curl -s https://aicomp.ai/v1/models -H "Authorization: Bearer $GATEWAY_KEY" \
  | jq -r '.data[].id'

Anything you request that is not on that list returns model_not_found, and vendor APIs do not fuzzy-match, so a near miss is a failure rather than a substitution.

What is available under one key. This is the part that changes what you can build. Instead of one vendor's catalogue, the same credential reaches models from several vendors, so a fallback or a per-task model choice becomes a string change rather than a second integration.

Where the bill comes from. Spend that was spread across vendor dashboards becomes one line, attributable per model.

What does not change

Worth being explicit, because migrations are often oversold:

• Your client code and SDK calls.
• Message formatting, tool calling shape, streaming behaviour (same protocol).
• Your prompts. They are model-specific, not endpoint-specific.

Prove traffic actually moved

This is the step people skip, and it is the only one that catches the dangerous failure. If the base_url override does not take effect. An environment variable the process never saw, a config that reads at startup, a default that wins (requests still succeed). They just go to the vendor. Tasks complete, nothing errors, and the migration silently did not happen.

# Send one deliberate request, then confirm it arrived
curl -s https://aicomp.ai/v1/chat/completions \
  -H "Authorization: Bearer $GATEWAY_KEY" -H "Content-Type: application/json" \
  -d '{"model":"gpt-5","messages":[{"role":"user","content":"ping"}],"max_tokens":8}'

Then check the endpoint's request log for that call. A 200 response proves the key works; the log entry proves the traffic is going where you think. Those are different claims and only the second one is the migration.

Rollback

Keep the old configuration in place rather than deleting it. Because the change is two fields, rollback is restoring them, no code path to unwind, no data to migrate back. That is a property of this specific migration and worth using: run both for a day, compare, then remove the old one.

What it costs once you are there

A day of ordinary application traffic — a few hundred thousand input tokens and a smaller volume of output: A day of work on this page's workload is 200k input and 50k output tokens — 4M input and 1M output across a 20-day month.

One day = 200k input + 50k output on this page's workload. 20 days = 4M input and 1M output. Rates checked 2026-10-27 (gateway rates move with upstream promotions), so verify the current number in your dashboard before committing to a budget.

The interesting column is not the cheapest row, it is that the same credential can reach all of them, so per-task model choice stops being an integration project.

Get API access

FAQ

FAQ

Related

• All tools with one endpoint — the same override in editors and CLIs
• OpenAI SDK with a custom endpoint — SDK-level detail
• Model fallback chain — what one endpoint makes easier
• OpenAI to Claude migration — when the model changes too
• How the gateway routes requests — what happens between call and model
• Model price index — every model with a page

Get API access