Models and inference

Use the access token saved during Sign in to populate a model picker and complete the first inference request.

1. List available models

Before showing a model picker, request the selected ChatGPT account’s model catalog with the same access token you will use for inference:

curl --fail-with-body --silent --show-error \
  https://api.openai.com/v1/models \
  -H "Authorization: Bearer $ACCESS_TOKEN" \
  | jq '[.models[] | select(.visibility == "list") | {slug, display_name}]'

The model-list response contains a models array. This example keeps models intended for display (visibility: "list") and preserves the server’s ordering. Show display_name in your UI and pass the selected slug as model in the next step. Refresh these choices when the user switches ChatGPT accounts. If you use Codex app-server, its model/list RPC may use a bundled or cached catalog. Use the request above when your UI needs current account-specific model choices.

2. Call the Responses API

To perform inference on behalf of the user, send the OAuth access token as Authorization: Bearer <ACCESS_TOKEN> to POST https://api.openai.com/v1/responses.

Supported endpoint: Use the public Responses API endpoint above for this ChatGPT plan usage flow; do not point it at ChatGPT’s backend-api endpoints. Use a model available to the signed-in account.

Set store to false and stream to true on each HTTP inference request in this flow. Consume the response as a stream of events.

Read the saved access_token for the selected ChatGPT account into the process environment as ACCESS_TOKEN before running these examples.

curl example

curl --no-buffer https://api.openai.com/v1/responses \
  -H "Authorization: Bearer ${ACCESS_TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6.1-sol",
    "input": [
      {
        "role": "user",
        "content": "Say exactly: Hello, world!"
      }
    ],
    "store": false,
    "stream": true
  }'

Python SDK example

Set the api_key parameter to the OAuth access token; the SDK sends it as the bearer credential.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30import os

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["ACCESS_TOKEN"],  # OAuth token sent as Bearer.
    base_url="https://api.openai.com/v1",
    max_retries=0,
)

completed = False
with client.responses.create(
    model="gpt-6.1-sol",
    input=[{"role": "user", "content": "Say exactly: Hello, world!"}],
    store=False,
    stream=True,
) as stream:
    for event in stream:
        if event.type == "response.output_text.delta":
            print(event.delta, end="", flush=True)
        elif event.type == "response.failed":
            error = event.response.error
            code = error.code if error else "unknown_error"
            raise RuntimeError(f"Response failed: {code}")
        elif event.type == "response.completed":
            completed = True

if not completed:
    raise RuntimeError("Stream ended without response.completed.")
print()

3. Wait for completed inference

Treat inference as successful only after receiving response.completed. Read through the terminal event. A usage-limit error can arrive as response.failed with subscription_sharing_usage_limit_exceeded or subscription_sharing_usage_unavailable after streaming has begun. Apply the corresponding recovery in Structured Responses errors. Handle response.incomplete, explicit errors, and interrupted streams separately from response.completed.