How to Use a Custom GPT in ChatGPT—and When to Use an API

How to use a Custom GPT

To chat with a GPT someone created, sign in to ChatGPT and open a GPT you have access to—either from a link they shared or from the GPTs area. Send it a prompt as you would in a regular ChatGPT conversation. You must be signed in to chat, and access can depend on the GPT’s sharing settings and your workspace permissions. (help.openai.com)

For example, try: “Summarize this report in five bullets. Separate facts from recommendations, and flag anything the report doesn’t answer.” State the task and format you want, then check important claims against reliable sources.

Retirement notice: OpenAI says Custom GPTs are scheduled to retire on December 11, 2026. The later date of February 11, 2027 applies to FedRAMP workspaces and certain Enterprise workspaces with an approved deferral—not every workspace. Existing GPTs remain usable until their applicable retirement date, subject to access and workspace permissions. Check the notice in your account or workspace and plan for a replacement if you rely on one. (help.openai.com)

Find a GPT you created or one shared with you

1. A GPT someone shared: Open the link they sent, or find the GPT in ChatGPT if it is available to you. Sign in if prompted, then start a conversation. If it does not open, check that you are using the right account or workspace and that the GPT is shared with you.
2. A GPT you created: Open it from the GPTs area or your available GPT list and start a chat. Access can depend on your account, workspace settings, and the GPT’s status during the retirement transition. (help.openai.com)
3. Write a specific first prompt: State your goal, relevant context, and desired format. For example: “Using the policy I provide, answer in plain English. Quote the relevant section, and say when the policy doesn’t cover the question.”

Can you create or edit a Custom GPT now?

OpenAI’s current rules distinguish creating a new GPT from editing an existing one. Personal ChatGPT accounts—including Free, Go, Plus, and Pro—cannot create or publish new GPTs. Editing an existing GPT may still be available if your subscription, account permissions, and workspace allow it. In Business, Enterprise, and Edu workspaces, check your workspace settings and notice. For affected Enterprise workspaces, new GPT creation ended on October 26, 2026; FedRAMP workspaces were excluded from that cutoff. An existing, unmigrated GPT can remain editable until its applicable retirement date; after migration, the original becomes read-only. GPT creation and editing are available on the web; mobile apps support using GPTs, not creating them. (help.openai.com)

Edit an existing GPT

If you have permission to edit it, use the web editor for the GPT you own. In the editor:

1. Change the fields relevant to your goal—for example, its instructions or reference material.
2. Test the revision in Preview with a normal prompt and one that checks a likely edge case.
3. Apply or save the revised version using the editor’s update/save control. You do not need to share the GPT to save an edit. The exact route and controls available can depend on your account or workspace; if the editor is unavailable, check its permissions and in-product notice. (help.openai.com)

Create a new GPT, if your workspace permits it

1. On the web, open Explore GPTs in the ChatGPT sidebar or go to the GPTs area, then select Create.
2. Use the conversational builder or configure the available fields directly.
3. Define the GPT’s task, intended audience, boundaries, and desired response format.
4. Add reference material or connected capabilities only if the use case needs them. For example, a handbook might be useful reference material for a policy Q&A GPT. Review data-handling and sharing controls before adding sensitive or restricted information.
5. Test in Preview, revise instructions when a test reveals a gap, and test again.
6. Save or update the GPT. Sharing is a separate, optional step: if sharing is available, open the editor’s Share controls, choose an available sharing level and permission, and save those sharing changes. Options depend on your account and workspace. (help.openai.com)

Test a GPT before relying on it

Use a small, repeatable check rather than judging it from one impressive answer:

• Normal request: Does it address the task you intended?
• Format check: Does it follow the requested length, structure, and tone?
• Missing-information check: If a necessary detail is absent, does it acknowledge the gap instead of inventing one?
• Source check: Can you verify important factual claims against the relevant document or another reliable source?
• Boundary check: Does it stay within the limits you gave it?

When a response misses the mark, revise the relevant instruction or example, then rerun the same prompt and your other checks. Testing helps you inspect behavior; it does not guarantee accuracy.

Custom GPT or production model API?

Use a Custom GPT to chat with a configured assistant inside ChatGPT. Use a model API when you are building an assistant experience into your own application. OpenAI’s documentation describes GPT Actions as connections a GPT uses to call external services. That is different from calling a saved ChatGPT GPT from your application. The official documentation reviewed here does not establish whether a saved GPT can be invoked directly through an API, so do not assume its instructions, files, sharing permissions, or Actions automatically carry over to an API integration. Treat direct invocation as unconfirmed unless OpenAI documents support for your use case. (developers.openai.com)

For an application, OpenAI documents the Responses API as a way to call models. You would configure the instructions, reference material, tools, and access controls your application needs; this is not a call to a saved GPT. Keep the API key on your server, not in browser or mobile-app code. (developers.openai.com)

A minimal server-side JavaScript example using the OpenAI SDK:

npm install openai

Set OPENAI_API_KEY and OPENAI_MODEL in your server environment, using a model available to your API project. Then run this in a Node.js file:

import OpenAI from "openai";

const client = new OpenAI();
const response = await client.responses.create({
  model: process.env.OPENAI_MODEL,
  instructions: "Answer clearly. If required information is missing, say so.",
  input: "Give me a three-bullet summary of why API keys should stay server-side."
});

if (!response.output_text) {
  throw new Error("The API returned no text output.");
}
console.log(response.output_text);

This sends input to a model and prints its text response; it does not invoke a saved Custom GPT. Check the current API quickstart and model documentation for setup and a model available to your project. (developers.openai.com)

If you are considering a GPT Action, keep the direction clear: an Action lets a GPT call an external API using configured authentication and an OpenAPI schema. It does not, by itself, establish that your application can call that GPT as an API. OpenAI also says Action availability can depend on workspace settings and supported models. (help.openai.com)

Sources

GPTs in ChatGPT

Creating and editing GPTs

Sharing and publishing GPTs

Custom GPT retirement and migration FAQ

Configuring actions in GPTs

GPT Actions

Developer quickstart

Responses Overview

Create a model response