Error messages

Rate limited by model provider, retrying…

Rate limited by model provider, retrying…

This shows in the agent's status line while a chat is waiting on a model:

Rate limited by model provider, retrying…

If the retries run out you get the real error. On September 3, 2026 a user on Cursor 3.18.25 saw "Rate limited" and then "We're having trouble connecting to the model provider. This might be temporary - please try again in a moment." A September 22 report pasted the JSON behind that second line:

{"error":"ERROR_NETWORK_ERROR","details":{"title":"Network Error","detail":"We're having trouble connecting to the model provider. This might be temporary - please try again in a moment.","isRetryable":true,"additionalInfo":{},"buttons":[],"planChoices":[]},"isExpected":true}

What "model provider" means here

Cursor doesn't run most of the models in its picker. The provider is the company behind the model (xAI for Grok, Anthropic for Claude, OpenAI for GPT), and it has its own capacity. When that company says "too many requests" (an HTTP 429, the standard status code for rate limiting), Cursor shows this line and tries again.

A Cursor staffer explained it on August 26, 2026, when Grok 4.6 was busy: the model "is temporarily hitting capacity, and Cursor is automatically retrying the request. It usually clears up on its own in a few minutes." It isn't your plan's usage running out. That's a different message, "You've hit your usage limit".

Why Auto doesn't always dodge it

We assumed Auto (Cursor's setting that picks a model for you) would quietly route around a busy provider. Mostly it does, but not in the first minutes of an outage. On September 3 the provider serving Grok went down from about 13:23 UTC, and staff wrote that "Auto picked a Grok model for your requests, so every retry hit the same unavailable endpoint." Auto chooses a model per request and "retries that same model a few times before failing". It shifts traffic when a provider is unhealthy, but a sudden outage can still catch you. The same day Anthropic had uptime problems, so Opus users saw it too.

• Pick a different model by hand in the model picker, ideally from another company. On September 3, GPT-5.6 worked while Grok was down.
• Or wait a few minutes and send again. Check status.cursor.com if it drags on.
• If it keeps failing, copy the Request ID from the chat menu and post it on the Cursor forum.

When it's your own API key

Cursor lets you add personal keys for OpenAI, Anthropic, Google, Azure and AWS Bedrock in Settings > Models. Requests for those models then go out on your key, and Cursor's docs say they "fail until you update or remove the key" if the provider rejects it. The limit in that case is the one on your provider account. The line looks like this:

User Provided API Key Rate Limit Exceeded. User API Key Rate limit exceeded

In May 2026 staff traced one of these to an Anthropic key capped at "30,000 input tokens per minute for your org", adding that "Cursor is just passing through the provider error." A Cursor developer said back in January 2025 that Cursor doesn't put its own limits on your key. The fix is a higher tier with the provider, or turning the key off so Cursor's plan serves the model.

The strangest case was on September 22. An Ultra subscriber titled their thread "Rate limited by model provider" and by their account every premium model had been failing for two months. Staff found an Azure OpenAI key toggle switched on with an empty Base URL. With any personal key toggle on, Cursor sends every model except its own (Auto, Composer, Cursor Grok 4.6) through that key, so only those kept working. Turning the toggles off fixed it straight away.

The older wording from 2025

Cursor 1.7.44 showed this in October 2025, sometimes to people already on Auto:

We've hit a rate limit with the provider. Please switch to the 'auto' model, another model, or try again in a few moments.

The error code behind it was ERROR_OPENAI_RATE_LIMIT_EXCEEDED. Two threads from October 14 ran to nearly 40 posts between them, and staff said the next day it "should have been fixed." If a model is simply swamped, Cursor may say High Load instead.