Help....Taking longer than expected

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

Extreme latency with Claude 4.5 Sonnet. For the past 2 days, every single request or “move to the next step” takes consistently over 2 minutes. A task with 6 sequential steps now takes 12+ minutes, making the product unusable for development.

Steps to Reproduce

Open any project (even a completely empty folder).

Start a chat or use Composer with Claude 4.5 Sonnet.

Ask for any code modification.

Wait for the response (it hangs for 2+ minutes).

Click “Move to next step” (it hangs for another 2+ minutes).

Operating System

Windows 10/11

Version Information

Cursor Version: 2.3.35 (system setup)
VSCode Version: 1.105.1
Commit: cf8353edc265f5e46b798bfb276861d0bf3bf120

For AI issues: which model did you use?

Claude 4.5 Sonnet

For AI issues: add Request ID with privacy disabled

82be4b1f-c4f2-4eb8-9819-25a51e905769
9697fc77-9b7c-477e-b0df-d5ad8ee75665

Additional Information

CRITICAL: > 1. Previously, the latest version gave “Taking longer than expected” errors. After I downgraded to 2.3.35, the error message disappeared, but the 2-minute latency remains (silent freezing).
2. I have already tried switching between HTTP/1.1 and HTTP/2 in the settings; it did NOT help.
3. This happens even in empty workspaces, suggesting a backend/routing issue.

Does this stop you from using Cursor

Yes - Cursor is unusable

Hey, thanks for the report and the request IDs.

I checked both request IDs, and the latency you’re seeing is caused by slow pool routing. This happens when your included Pro plan usage ((20) USD/month of API costs) has been exceeded. Each model call gets an extra ~90-second wait, and that adds up across multiple agent steps. With 6+ steps, it matches the 12+ minutes you’re describing.

Claude 4.5 Sonnet, especially in thinking mode, is one of the more expensive models, so it can burn through the included budget pretty quickly with heavy use.

A few options to fix this:

  • Enable usage-based pricing: This removes the slow pool throttling and charges API rates for anything over the included (20) USD. You can do this from your dashboard: https://cursor.com/dashboard
  • Upgrade your plan: Pro Plus ((60) USD/month) includes (70) USD of API usage. Ultra ((200) USD/month) includes (400) USD.
  • Switch to a less expensive model: Models like Auto or lighter models use less of your budget per request.

You can check your current usage breakdown at https://cursor.com/dashboard under the Usage tab.

This isn’t a backend outage or a routing bug, it’s just the usage limit kicking in. Let me know if you have questions about pricing.

Is this a new hidden rule implemented in the past few days? This behavior is honestly appalling.

My typical workflow requires 10-20 ‘Move to next step’ iterations per conversation. With this current latency, I am forced to wait an extra 900 to 1800 seconds (15-30 minutes) for a single task that should take only 1 minute.

Have you considered the user experience at all? This is a complete betrayal of the ‘productivity’ promise Cursor stands for. Since the product has become functionally useless for my professional work, I will be applying to terminate my annual subscription and requesting a refund. This is not the service I paid for.

This isn’t a new or hidden rule. The slow pool has always been part of how included usage works on all plans. The Pro plan includes (20) USD/month of API costs, and Claude 4.5 Sonnet, especially in thinking mode, is one of the most expensive models, so heavy use can use up that budget fast.

A few things that would immediately fix the latency:

  • Enable on-demand usage at https://cursor.com/dashboard. This removes the slow pool entirely and charges you API rates for anything over (20) USD. You stay on Pro, no plan change needed.
  • Switch to a cheaper model like Auto. It routes to the best model for the task and costs a lot less per request.
  • Upgrade. Pro Plus ((60) USD/mo) includes (70) USD of API usage, Ultra ((200) USD/mo) includes (400) USD.

You can see exactly where your budget went under the Usage tab on your dashboard.

But gemini-3.0-flash also needs 90 second waiting each step!!!

25808932-72f8-42bd-966c-9d0ea4a21a18
can you help check why this request always taking longer than expected though it still in limitation?

Network Diagnostics show that Agent Client network socket disconnected before TLS connection established.

This situation often happens for AUTO mode. It little happen when using GPT-5.4. Is that something happen in your server?

This seems to be happening a lot in the past couple weeks. No matter what model is chosen, it runs extremely slow, losing so much productivity. I’m getting a lot of requests that finish but show as 0 tokens in the usage dashboard. Not all the slow requests are like that, but that could be a related issue.

This topic was automatically closed 22 days after the last reply. New replies are no longer allowed.