[Regression] Cursor Grok 4.6 High quality dropped sharply this morning — worse than last week’s Grok 4.5

Support Help
1

Summary

This morning (2026-09-20, UTC+8) I noticed a sharp quality drop on Cursor Grok 4.6 at effort = High. Output is consistently weaker than Grok 4.5 (High) from the past week — not a single bad chat. Please check whether routing, serving, system prompt, or effort mapping changed.

Environment

  • Cursor version: [Help → About Cursor → 粘贴完整版本]
  • OS: Windows 10 / 11, build 26100
  • Model: Cursor Grok 4.6
  • Effort: High
  • Fast: [on / off]
  • Privacy: [Privacy Mode / Share Data]
  • Plan: [Pro / Pro+ / Ultra]

Request IDs

Please look up these requests (chat ⋯ → Copy Request ID):

  • Grok 4.6 High (this morning, bad): [paste]
  • Same task on Grok 4.5 High (baseline, better): [paste, if available]

If Privacy Mode was on for the original runs, I can temporarily enable Share Data, reproduce, and post new IDs.

What changed

  • Baseline: last week, Grok 4.5 High was usable for my usual Agent work (instruction following, planning, edits).
  • Today: Grok 4.6 High is markedly worse — lower-quality reasoning and deliverables, not just slower.
  • Same machine, same project type, same workflow. I did not change rules/prompts in a way that would explain this.

Expected vs actual

  • Expected: Grok 4.6 High ≥ last week’s Grok 4.5 High on instruction following and Agent task quality (as documented).
  • Actual: 4.6 High is clearly below that baseline this morning.

Concrete example

[用同一任务各跑一次 4.6 High 和 4.5 High，各写 3–5 句：你要求什么 / 期望 / 实际差在哪。不要贴密钥或私有代码。]

What I already tried

  • New chat
  • Compared the same prompt on Grok 4.5 High
  • Confirmed model + effort in the picker (not Auto)

Ask

Please confirm whether anything changed in the last 24–72 hours for cursor-grok-4.6-high (routing, provider, prompt, effort). Happy to add more Request IDs after a Share Data repro.

6

Hey @LiMaoHua,

Quick thing to try first: Grok 4.6 “High” and last week’s Grok 4.5 “High” aren’t the same reasoning level - 4.6 has an extra tier above High. For a like-for-like match to the 4.5 High you were on, switch to Cursor Grok 4.6 at Extra High, in a fresh chat (long chats get summarized and quality drops).

Nothing changed on our side for Grok 4.6 High in the last few days (routing, effort, and prompt are all the same), so that effort difference is the most likely cause.

If it still looks worse after that, reply with the Request ID of a weak run (chat ⋯ → Copy Request ID) and a line on what fell short, and I’ll dig in.

7

Thanks for your reply. I’m still trying on my end. I’ve found that grok4.6 high fast is indeed slightly less capable than it was last week—it often can’t understand the meaning of tasks it could easily understand last week. Could this be due to an issue with how your tokenizer algorithm adapts to Chinese?

8

Hey @LiMaoHua, nothing changed on our side for Grok 4.6 High in the last few days: same model, reasoning level, prompt, and routing. Two things make it feel weaker than last week’s 4.5 High:

  1. They run at different reasoning levels. For a like-for-like test, use Grok 4.6 at Extra High.
  2. Long chats get summarized and quality drops. Start a fresh chat per task.

You’re on Nightly (3.21.16), which changes daily, so try the same task on Stable too. If it’s still worse, send a Request ID of a bad run (chat ⋯ → Copy Request ID) and one line on where it fell short.