On-Demand Usage OFF blocks OpenAI-compatible / Vercel AI Gateway models (looks like misconfig)

[ta9to]

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

I configured Vercel AI Gateway in Cursor as an OpenAI-compatible (BYOK-style) provider: API Key, Base URL, and Model ID were set. Every model selection still failed with errors that looked like a Gateway / provider misconfiguration.

I double-checked Vercel AI Gateway settings and available models; the Gateway side looked fine, and requests still failed in Cursor.

Eventually I noticed On-Demand Usage was OFF in Cursor. Immediately after turning On-Demand Usage ON, the same Vercel AI Gateway models started working with no other config changes.

The confusing part: with an external / BYOK-style setup, Cursor’s On-Demand Usage setting can block requests in a way that looks like a Vercel AI Gateway misconfiguration, not a Cursor billing gate.

Steps to Reproduce

1. Configure Cursor Models with an OpenAI-compatible provider pointing at Vercel AI Gateway (API Key, Base URL, Model ID).
2. Ensure On-Demand Usage is OFF in Cursor account / spending settings.
3. Select a Gateway-backed model and send a chat / agent request.
4. Observe errors (while Gateway settings/models appear correct).
5. Turn On-Demand Usage ON.
6. Retry the same model / request with no other changes.

Expected Behavior

If Cursor is using an external OpenAI-compatible / BYOK provider (e.g. Vercel AI Gateway) with a valid key and base URL, requests should work (or fail with a clear message that a Cursor billing setting is blocking them)—not with errors that read like the external API is misconfigured.

Operating System

Other

Version Information

Not provided

Additional Information

Actual Behavior

With On-Demand Usage OFF, Gateway-backed models failed in a way that suggested a Vercel AI Gateway / OpenAI-compatible setup problem. Turning On-Demand Usage ON made the same configuration work immediately.

Workaround

Enable On-Demand Usage in Cursor. After that, the previously failing Vercel AI Gateway models worked without changing API Key / Base URL / Model ID.

Open question

Is requiring On-Demand Usage for external OpenAI-compatible / BYOK providers (including Vercel AI Gateway) intended product behavior, or a bug / UX issue (billing gate presented as a provider error)?

Related

• Forum (BYOK blocked / On-Demand workaround): Get more usage to switch models?
• Docs (On-Demand / usage-based): Usage-based charges | Cursor Docs

Also related by title only (no URLs, due to new-user link limit): “On demande usage limit even when using my own API keys”; “AWS Bedrock + Auto/Premium — Which quota is actually being used?”; Models & Pricing and Team Pricing docs (BYOK / Cursor Token Rate).

(Observed behavior only; not asserting root cause or product intent.)

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

[Colin]

Hey there.

Have a look here.

I use my own API key, but the actual usage is still counted under "Other Models." As a result, when the usage limit for "Other Models" is reached, I can no longer use my own API key

Hey there! This is actually expected behavior on Teams and Enterprise plans, and I can see why the docs made it look otherwise. Here’s what’s going on: On Teams/Enterprise plans, third-party models carry a Cursor Token Rate of $0.25 per million tokens, and that applies to models you run with your own API key too (BYOK or not). The actual model cost still goes to your own provider account, but the Cursor Token Rate is real usage, so your own-key requests appear under Other Models and draw from …