# Cursor and Your ChatGPT Subscription: What Works in 2026

Cursor's OpenAI key override can point chat at a subscription-backed endpoint. What survives the override, what stays on Cursor's stack, and when it pays off.

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

Cursor cannot sign in with a ChatGPT account, but it accepts a custom OpenAI API key and a base URL override. Point both at `https://api.proxyllm.ai/v1` with a ProxyLLM key and Cursor's chat requests to OpenAI models run through [Codex Hosted](/) on your own ChatGPT subscription instead of metered API billing. The override covers chat-style traffic; Tab autocomplete and Cursor's native features stay on Cursor's own models and its own plan.

That second sentence is the honest part most write-ups skip, so this article covers both halves: the setup and the boundaries.

## Cursor has two auth stories

Cursor's own subscription signs you into Cursor's product: Tab completion on its custom models, its hosted frontier-model access, its agent infrastructure. Separately, Cursor's settings accept your own OpenAI API key for model traffic you want billed to you, and that key field comes with a base URL override for OpenAI-compatible endpoints.

The override is the opening. Anything that accepts an OpenAI base URL can point at a subscription-backed endpoint, and Cursor's custom-key mode qualifies. Your ChatGPT plan does not plug into Cursor directly; it backs the endpoint Cursor talks to.

## Configure the override

In Cursor's settings, under the model and API key section:

```text
OpenAI API Key:            pk_live_your_proxyllm_key
Override OpenAI Base URL:  https://api.proxyllm.ai/v1
```

Paste the ProxyLLM key, enable the base URL override, set the URL, and let Cursor verify. The verify step performs a real request, so a pass means the lane works end to end. Behind the endpoint, your ChatGPT account connects once: you run OpenAI's login command on your machine, then one connect command; the [setup guide](/blog/codex-hosted-setup-guide) covers those five minutes.

Give Cursor its own scoped key rather than reusing one from another tool. Editor traffic is bursty, and a per-key budget cap plus the request log tells you exactly what the editor sent and which lane served it.

## What survives the override

The honest capability table, as of June 2026:

| Cursor feature                         | With a custom key and base URL                    |
| -------------------------------------- | ------------------------------------------------- |
| Chat with OpenAI models                | Works; billed through your endpoint               |
| Tab autocomplete                       | Unaffected; Cursor's own models, Cursor's billing |
| Agent features on Cursor-served models | Stay on Cursor's plan                             |
| Features Cursor restricts to its keys  | Vary by version; some reject custom keys          |

Two caveats worth stating plainly. Cursor ships fast, and which features accept custom keys shifts between releases, so treat the table as a snapshot. And responses on the Codex lane arrive complete rather than streamed: in-editor chat shows a wait, then the full answer, instead of text typing out. Some developers do not notice; some find it jarring. Decide after a day of use, not from this paragraph.

If the override feels too narrow for how you work, a harness built around OpenAI-compatible providers gives the subscription lane more surface area; [opencode is the cleanest example](/blog/opencode-chatgpt-subscription).

## When the override actually pays

Run the numbers before bothering. A heavy chat user sending 80 requests a day at roughly 6,000 input and 700 output tokens each, on GPT-5.4 (OpenAI's June 2026 list: $2.50 per million input, $15 per million output):

```text
Input:  80 × 6,000 × 30 = 14.4M tokens × $2.50/M = $36.00
Output: 80 × 700 × 30   =  1.7M tokens × $15/M   = $25.20
Editor chat alone                                ≈ $61/mo
```

Sixty-one dollars of metered chat does not justify a $129 platform fee plus a $20 plan on its own. One editor's chat traffic rarely justifies a gateway; a whole toolbelt on one flat plan does. The math works when the same endpoint also carries your coding-agent sessions, scripts, CI jobs, and automation, pooling onto one ChatGPT plan with one request log. On our planning estimates a Plus plan absorbs roughly $700 of API-equivalent work a month, which is a lot of room once several tools share it.

The boundary cases are covered in [what works with Codex Hosted](/blog/codex-hosted-for-openai-compatible-apps): chat-model traffic fits, embeddings and fine-tunes stay on a real API key.

## The policy posture, briefly

Programmatic Codex use backed by your own ChatGPT account is documented OpenAI functionality, you sign in with OpenAI's own login command on your machine, and we never see the password. One account runs in one isolated container, never shared or pooled. That is the intended functionality, and OpenAI has the final call on its services.

The condensed setup lives on [the Cursor integration page](/integrations/cursor). If you want to see whether your combined tooling spend clears the breakeven, the [calculator](/calculator) answers it with your own numbers.