# Use Your ChatGPT Subscription in n8n (No API Key Billing)

Point n8n's OpenAI credential at a Codex-backed base URL and your workflow's AI calls bill to your flat ChatGPT plan. Setup steps, caveats, and the cost math.

*Published 2026-06-12 · https://proxyllm.ai/blog/use-chatgpt-subscription-with-n8n*

You cannot type a ChatGPT login into n8n, but you can point n8n's OpenAI credential at an endpoint backed by your ChatGPT plan. n8n's OpenAI credential has a Base URL field; set it to a Codex-backed endpoint like `https://api.proxyllm.ai/v1`, and every OpenAI node and AI Agent in your workflows bills against your flat subscription instead of a per-token API key. Setup is one credential edit.

Here is why the direct route fails, the exact configuration, and what a real workflow costs both ways.

## Why your ChatGPT plan does not plug in directly

This is the most common confusion in n8n forums: people pay $20 a month for ChatGPT, wire up an OpenAI node, and get billed separately per token. ChatGPT and the OpenAI API are two different products with two different bills, and the plan includes no API credits. We wrote up the split in [why your OpenAI API bill is separate from ChatGPT](/blog/openai-api-billing-separate-from-chatgpt).

The bridge is Codex. OpenAI's coding agent is included in ChatGPT plans, signs in with your ChatGPT account, and runs programmatically through `codex exec`, its documented non-interactive mode. [Codex Hosted](/) runs that CLI for you in an isolated container, signed in by you (you run the login command on your own machine), and exposes it as an OpenAI-compatible endpoint. Programmatic Codex use is intended functionality, and OpenAI has the final call on its services; we keep one account in one container, never pooled.

If you have not connected an account yet, the [setup guide](/blog/codex-hosted-setup-guide) covers it: sign in on your machine, one connect command, container start, base URL. About five minutes.

## Configure the n8n credential

1. In n8n, open **Credentials**, then **Add credential**, and pick **OpenAI**.
2. Fill the fields:

```text
API Key:         your ProxyLLM key
Organization ID: leave empty
Base URL:        https://api.proxyllm.ai/v1
```

3. Save. The credential test performs a real request, so a green check means the lane works end to end.

That is the whole integration. Any node that accepts an OpenAI credential now routes through your subscription: the **OpenAI** node, the **OpenAI Chat Model** attached to **AI Agent** or **Basic LLM Chain** nodes, and anything else built on the same credential type. If a model dropdown does not populate, enter the model ID directly, for example `gpt-5`. You can also skip credentials entirely and call the endpoint from an **HTTP Request** node with an `Authorization: Bearer` header; the [n8n integration page](/integrations/n8n) shows both patterns.

## Which nodes work, and the two caveats

Chat-completion traffic is the sweet spot, and it is most of what n8n workflows do. Two caveats, stated plainly:

**Responses arrive complete, not streamed.** The Codex lane returns the full answer in one response. For workflows this is invisible: a workflow node does not need streaming, it needs the complete answer before the next node runs. Only if you build a user-facing streaming chat on top of n8n should you keep an API-key lane for that path.

**Embeddings stay on a real API key.** The model surface is what Codex serves, which means chat models. Point embeddings nodes, and anything like fine-tuning jobs, at your own OpenAI key. ProxyLLM passes BYO keys through with no markup, so both lanes can live behind the same endpoint.

## What a real workflow costs

Each AI Agent node turns one execution into several model calls, and the meter bills every one of them. That multiplication is why n8n workloads outgrow per-token billing fast.

Worked example: a support-triage workflow where an AI Agent classifies, looks up context with tools, and drafts a reply. It averages 5 model calls per execution at roughly 3,000 input and 800 output tokens per call, and runs 1,000 executions a day.

| Monthly figure        | Amount                                            |
| --------------------- | ------------------------------------------------- |
| Model calls           | 1,000 × 5 × 30 = 150,000                          |
| Input tokens          | 450M, at $2.50/M (GPT-5.4) = $1,125               |
| Output tokens         | 120M, at $15/M (GPT-5.4) = $1,800                 |
| **API total**         | **$2,925/mo**                                     |
| **Subscription path** | **ChatGPT Pro 5x $100 + ProxyLLM $129/mo** |

Token prices are OpenAI's June 2026 list for GPT-5.4. The Pro 5x window covers roughly $3,500 of API-equivalent work on our planning estimates; treat that as an estimate, not a guarantee, and watch the per-lane request log for your real consumption. If the workflow bursts past a window, requests fall back to a second account or your own API key until the window resets.

The same swap works in Make and Zapier with their own mechanics; the [Make version of this guide](/blog/use-chatgpt-subscription-with-make) is next if you run both platforms.

Run your own numbers against your current OpenAI bill in the [calculator](/calculator); if the workflow above looks like yours, the math takes thirty seconds to check.