Migrate from OpenAI

Create a key and add credit

Sign up at app.glexir.com, create an API key on the API keys page, and add credit on the Billing page. Glexir is prepaid: requests are refused with 402 credit_balance_exhausted when your balance can't cover them.

Usage is billed at OpenAI's list price. Larger single top-ups add bonus credit. See Pricing for the tiers and Credits & billing for how balances work.

Create one key per environment, for example staging and production. See Best practices.

Check that your models are sold

Glexir sells text, embedding and speech-to-text models that have a price in its price book. List them:

import Glexir from "@glexir/sdk";

const glexir = new Glexir({ apiKey: process.env.GLEXIR_API_KEY });

for await (const model of glexir.models.list()) {
  console.log(model.id);
}

Use the model alias (gpt-5.4-mini), not a dated snapshot name (gpt-5.4-mini-2026-03-17). Snapshot names return 403 model_not_priced. Responses may still report the snapshot name in model; that is expected and is billed at the alias's price.

Point your client at Glexir

Replace your OpenAI key with your Glexir key. @glexir/sdk calls https://api.glexir.com/v1 by default. With the openai package, also set the base URL to https://api.glexir.com/v1.

Switching to @glexir/sdk is a small change, because its resources and method names match the OpenAI SDK for every endpoint Glexir supports:

@glexir/sdk sends an Idempotency-Key on every non-GET request, so its automatic retries are never billed twice. With the openai package you send that header yourself.

Remove unsupported features

Glexir rejects some OpenAI features before the request goes upstream. You are not charged for rejected requests. The full list is in What Glexir rejects. The usual changes:

• Hosted tools. Replace web search, file search, code interpreter, MCP and similar tools with function or custom tools that your code runs.
• Prompt templates. Replace prompt: { id: "pmpt_..." } with the text in instructions.
• Audio in chat. Send audio to /v1/audio/transcriptions first, then pass the text to the model.

Recreate stored state

Objects you created on OpenAI do not exist on Glexir. Requests that reference them return 404.

• Files. Upload them again through Glexir, with purpose: "user_data" (model input) or purpose: "batch".
• Stored responses and conversations. Start new ones. previous_response_id and conversation only accept IDs created through Glexir.
• Stored chat completions. Only completions created through Glexir with store: true are listed.
• Batches. Submit new batches from input files uploaded through Glexir.

If you keep conversation state yourself (sending the full history, or store: false and replaying output items), nothing changes.

Handle Glexir-specific errors

Add handling for the statuses OpenAI does not return in the same way:

• 402 credit_balance_exhausted or spend_limit_exceeded: alert someone and stop retrying.
• 429 concurrency_limit_exceeded: limit how many requests you run at once.
• 409 and 422 with idempotency_* codes: see Idempotency.

See HTTP error codes.

Test and switch over

Run your test suite against Glexir with a staging key. Compare outputs and token usage with your OpenAI baseline, then switch production. Watch the Usage page in the dashboard for the first days.