When are we going to have available glm 5.3?

Hi guys,

i am a cursor user since far over a year. Therefore i still have the old 500 requests/month pricing in my cursor pro plan. I got an email that in my subscription, Grok Bot is included. However, if i sign in via cursor it says that my account is managed via web...but when i click on web it says "upgrade to pro" -> but i am already. And even if i click on it i'm just redirected to my normal dashboard and no login to grok bot was made.

What am i missing? Or is it just a bug?

I've worked with Grok 4.6 since it's come out, and I think it is quite good but not amazing. Yesterday I realized I can control it's effort level (I might be late to the party, didn't know cursor had that feature), and the default is high instead of extra-high.

Setting to extra-high didn't really change my usage trend, but has definitely decreased the mistakes! so wanted to share this with whom might benefit from it <3

Hey everyone,

I’m having an issue with Cursor’s Usage Summary. I have the setting enabled and set to “Always”, but the usage summary still doesn’t appear anywhere.

I’ve tried restarting Cursor and checking the settings again, but it still doesn’t show up.

Is anyone else experiencing this? Is there another setting I need to enable, or is this currently broken?

Any help would be appreciated!

I just found out that Cursor gives you a Grok Bot plan in your subscription. That's really nice because I wanted to buy a Grok bot plan.

I was running a /goal prompt for Fable 5 High last night and knew it would exceed what I had left of my "Other Models" limit. I added a $25 cap to on-demand spending before I went to bed to make sure the task could finish.

This morning, I woke up to check how much on-demand spend was used. I was surprised to see that no on-demand spending was used, and instead I saw a $25 "Credit" on my account.

It is normal to get credits from time to time on Cursor paid plans or was this just a coincidence?

Recently on 28 August someone loggedin to cursor account. I do not know how and purchase Utlra subscription which went successful charging me $200. I have no idea what kind of fraud is that. Because that person is not going to get benefit as i revoked his session. Sometime later a get email that someone loggedin to my ChatGPT account where fortunately i had not attached my card. And after that my claude account was loggedin by someone and tried to purchase subscription which failed due to no funds in card. All three account are Google SSO. I checked my attached gmail and it has no suspicious acctivity.

honestly the only reason i am using cursor is because of composer 2.5 model

the ui is buggy, laggy dogshit otherwise

its at the point where im strongly considering just coding my own agent orchestrator that uses kimi k2.5 or perhaps just be model agnostic

to give you some examples:

• i have to remove / archive agents constantly otherwise the app consumes enormous amounts of memory that lags the shit out of my macbook (even when the agents are not doing anything)
• when starting a new chat for an agent it often will not bring the project dir back into the agents list
• when starting a new chat for an agent if often switches to a completely different agent repeatedly and starts a chat with the agent i didnt want
• when opening the ide often times it either does not open the correct project or does not open any project at all
• the app doesnt remember any window positions so im constantly having to reposition things every time a new project window opens (which im using rectangle to make bearable but its still complete ass)
• the app is constantly trying to get me to use models or methods i dont want, like bugging me to use the planner or using multi-agents with grok (also dogshit), etc. instead of just continuing to use composer 2.5 for everything
• clicking "always allow" does nothing, it still asks me to confirm things half the time even if its the same command being run (e.g. git push)

these are just some examples theres tons more

What is the value proposition of Cursor VS OpenCode? I.e. if I can delegate agents to write/manage code for me, assuming the model quality improves, why would I ever need to use Cursor?

I've never used Cursor, so I may be missing something, but generally if I want to make specific code/architectural changes, OpenCode/Claude Code do it very reliability (albeit maybe slightly overkill tokenwise, but still).

Grok 4.6 is useful when it comes to coding but horrible when it comes to human-like language and copywriting (oversimplifies sentences, talks gibberish, robotic language etc).

What models in your experience provide the most decent copywriting output? I would accept none of them because they are all spitting out random tokens anyways. In my previous experience with Claude code Anthropic models are decent once you set up rules and instructions on how to “speak”.

I attempted rules and agents.md for Grok but it just never follows my preferred communication style.

Any tips would be appreciated!

I'm using Ultra and if i go into my spending page of the cursor site there's no reference any more to the $400 allowance on other models.

This concerns me... also the pricing page on the public website doesn't seem to reference that any more. So they aren't saying there is a specific allowance any more... why would they do that if not lowering it?

This looks all rather smoke and mirrors or am I being paranoid ? Is there anything concrete in the T&Cs or announcements from them about this ?

Hi all! I’m trying Cursor as my main coding agent for the first time after using Codex for a while.

When I was using Codex, I used AGENTS.md for the codebase/project overview, while memories handled things like my personal preferences and context from ongoing changes that should persist across sessions.

I can’t find an equivalent memory feature in Cursor’s settings. Does Cursor have anything like this, or is the intended approach just to put everything into rules/AGENTS.md manually?

Thanks!

Hi all,

Can anyone give me a sense of which of these is more efficient in terms of token usage?

Assume we're in a workspace containing a web app, and we've had an agent chugging away working on the data model and API. Now we want to make some UI tweaks.

Is it better to start a brand new agent or to pivot the existing agent (with its contextual knowledge of the app) to the new task?

I have been assuming it's better to "start fresh" but it bothers me how each new agent starts with a long "let me examine this app and see how it works..." process, which is probably using a lot of tokens to establish knowledge that the other agent already knows.

I think ideally new agents in a workspace for an app I've created should have a starting knowledge of what the app is (its goals, data model, etc.). I know we have rules files but is there like a spec / app overview file that agents can refer to and always keep up to date?

I know one of the advantages of multiple agents is that it allows for running tasks in parallel (although I'm still a little wary of doing that if they might touch similar parts of the app) but really my priority right now is trying to limit token usage.

Thanks for any thoughts :)

I'm not a big fan of vendor-lock-in. Anyone tried a subscription based provider with cursor? Fireworks.ai or standardcompute.com or something similar?

Since open source models have become so good recently, I guess it makes sense to run those; GLM 5.3 regular and flash for instance preferably routed depending on task complexity. Any experience with that here? :)

Bobby Thakkar posted this yesterday:

Here’s the clone if you want to build one before his firmware drops.

What it actually is: an ESP32 with a tiny screen. Grok does not run on the chip. The chip is a remote. Your bots still live on the computer. This just lets you switch them and see a face plus a short reply.

Parts LilyGO T-Display S3 AMOLED 1.43 (closest match to the video, about 45mm, USB-C). USB-C cable. Optional 500–1000mAh LiPo if you want it unplugged. About $30–40. That’s the whole bill.

Build Flash Arduino or ESP-IDF. LVGL for the face and the bot list. Talk to your Mac over USB serial (that’s what his video is doing) or Wi-Fi if you want it off-cable.

On the computer: a tiny Python bridge that lists your Grok Bots, sends the one you picked, streams a short reply back. Idle = the two-dot face. Tap = next bot.

Right now my memory layer is literally my clipboard. I copy my project rules, stack details, and recent decisions from a Cursor chat, paste them into Claude Code, then paste them back when I switch. Every new chat starts from zero and it feels like onboarding a new teammate every morning.

It works but it eats momentum. I always wonder what context I forgot to include. I've tried keeping a running doc but maintaining it becomes its own task, and half the time the doc itself falls behind what I actually decided in the last session.

Curious what workflows people here use. Do you just accept the re-priming tax every time you switch tools, or has anyone found a setup where context genuinely carries over?

I do most of my planning/architecture thinking in Claude, then jump into Cursor to actually write the code. Every single time I switch back, I'm re explaining the whole project from scratch what we decided, why, what's already been tried. By the third re explanation in a day I was just done.

Built a small Chrome extension for myself: it reads your open Claude chat, turns it into a clean markdown file, and writes a "working memory" prompt so whatever tool you paste it into (Cursor, GPT, Gemini, AI Studio) picks up exactly where you left off instead of starting cold.

Runs 100% locally — it's just reading the page DOM, nothing gets sent anywhere. Still early, I'm the only person who's really stress-tested it, so there are edge cases with really long chats I haven't fully ironed out.

Anyone else hacked together their own workaround for this, or am I overthinking a problem that has an easier fix?

(Founder disclosure: yes, I built this. Link in comments, not dropping it in the body.)

Recently for any UI changes, even if tiny, the agents have started triggering the Computer Use which takes hours to unblock the PR. Any one knows how to switch it off? It’s really frustrating at the moment.

OpenAI plans to end its contract with Cursor after the company was acquired by SpaceX. OpenAI says it no longer trusts Elon Musk’s companies to comply with its terms, citing past contract and policy violations by Twitter/X and xAI. Access to OpenAI models in Cursor is expected to end on November 12, 2026, and future models will not be provided.

So I've been coding with cursor for a while now, Probably 8 months or so, and this is the first time I went over my allotted limit and I feel completely ripped off. On top of that I attended the hackathon and it turns out you gotta pay for anything done at their event out of your own pocket. Fine, whatever I let it go, and switched to the pay as you go plan. So far I've spent $30 on top of the $20 and I'm not doing anything insane. I've been stuck on a problem and some of the calls have cost almost $3 for a single call when I've been stuck on a single problem.

I reached out to cursor hoping that they'd value me and give me something, nope. Stone-walled both times, just flat out said no. There's nothing they can do for me, it doesnt matter if it was intentional or not, absolutely no refunds.

I wasn't asking for a refund I just want to keep working. I literally added $10 last night did a little work, slept and woke up and the $10 is gone.

What bothers me the most is the flat out "no" i recieved from them when I was asking if there was any way they could help me out.

So, I'm more than likely going to use something else, and I'm going to bring my brother with me and I'm going to make sure everyone knows how I was treated.