Hey - I see you’ve commented this on a few threads! We introduced caching to bedrock which should actually lead to overall lower costs per token.
Hey @andrewh - What does caching have to do with the issue we’re seeing? Can you clarify?
We want to use the 200k model, but Cursor keeps invoking the 1M model instead, which doubles the cost and has lower RPM and TPM quotas than the 200k model. So, caching is welcomed but doesn’t solve the bug you have, we still over pay and cannot use Cursor as intended according to your docs.
For reference, here is the issue I opened.
I responded in another Bedrock-related thread as well to check whether others see the same behavior.
Separately, if this ties into your recent changes to reduce consumption, I’ve noticed since yesterday that my chats are being summarized far too early, even when they’re well under 5% of the context window, and sometimes during the first agent response. That isn’t expected behavior and isn’t something we want. Screenshot below.
Completely understand and shared with the team - we’re looking into how to give you more optionality. However, caching introduced to Cursor’s Bedrock integration reduces costs more than the different context variants. Better caching via model API provides reduces costs significantly.
Adding a +1 here
Almost every day there is a new update popup. I’m getting afraid to click.
The last couple days have had some REALLY bad bugs.
Yesterday woke up to like 20 file deletions in staging. Today agent convos suddenly stop taking new responses.
Maybe the update popup could show what it covers so we can decide if we want it?
I’m installing them right now in the hopes they fix my other bugs, but it’s kinda russian roulette because they might contain new bugs that are worse.
But you remove features from newer releases that we depend on, so what are we supposed to do?
Before you push new features , you should test it at first.
C’mon guys, how many of you really are completely stuck because bugs? At least I can chat just fine with the models and they do their work mostly just fine. There might be room for improvement here and there, but mostly I just ignore those and continue my vibe coding
Literally my bug report above makes working through Cursor impossible for me.
Not sure what is the real problem there? But why not just keep another terminal (outside Cursor ide) open and start from there? I have to also launch Aspire project from visual studio if I want to have working breakpoints, but life is.
This really hits the nail on the head.
Like, I learnt to code 1-2 years ago, by myself, and my main focus is just C# for Unity.
And still I’ve kept seeing stuff for the cursor development that I’m just… what are you doing?
I sometimes feel like they are genuinely vibe coding the whole application and they have no idea or understanding about the actual code themselves, but that also feels to crazy to be true.
It took them months to give the ability to simple give the ability to accept auto-complete using a different button then TAB, even when they have a fork of VScode itself as the base.
Meanwhile I threw together a AutoHotkey script myself to work around it by actively rebinding and simulating keystrokes.
It’s like the changelogs, they keep refusing to provide proper and timely changelogs.
Why?
They keep refusing to provide any kind of LTS or stable branch.
Why?
They insist on pushing users towards new releases.
Why?
I’ve been using Cursor for a long time, and I swear I genuinely do not feel that the actual output quality has gotten significantly better than it was in 0.47 using Claude 3.7.
All that has happened since is a lot of random new fancy concepts and renaming, while the actual quality of use has not significantly changed or improved.
If focus had actually been put into just improving the core functions, I can only imagine how good a product they could possibly have had at this point given the crazy capital they have on their hands. They are so fat with cash it’s genuinely hard to wrap your head around, but it sure isn’t going into the product itself that I can say for sure.
- Because the Agent launches the terminal for self-checks and for manual testing on my end, not me.
- Because the Agent may need to launch the terminal to work and I don’t want it to remain hanging until the computer is restarted or until I find it in Task Manager.
- Because the Agent may accidentally or intentionally start a process in the terminal in such a way that the CLI interface will require additional user input, and I can’t respond in read-only mode.
- Because this feature worked in Cursor from the moment I started using it, and it’s one of the key tools of AI Agents.
Over the past 4 months, there has been only one change that I liked (changing the style of the diff code blocks in the Chat and the Editor), but every 2 weeks there is an update that makes me hate Cursor.
Edit: One could have mentioned the console improvements I liked and the addition of a browser, if the console improvements hadn’t been screwed over by the release of the new console, and if the browser hadn’t become a forced replacement for the web search tool that hasn’t been working for three months.
That’s like saying having to exit your game and manually copy the game directory to save your game is “just life”. It’s not, that’s your game that you paid/paying for that isn’t providing basic functionality that you are perfectly valid for expecting to function.
There is noting strange about expecting reasonable functionality and quality from a product you are paying for, nor is voicing your opinions or concerns over such.
If anything, the strange behavior is to actively promote acceptance of said deficiencies.
Why are you actively spending time trying to validate a lack of quality for a product you yourself use?
Are you saying that having to launch secondary applications is non-problematic or even preferable, and that anyone feeling otherwise is seeing a problem you don’t consider problematic - in fact, you consider the opinion of not having to launch a secondary application is more problematic?
Why?
Are you saying you have built your whole workflow on the ability of agent to perform same task twice in row 
You may look at https://github.com/Artemonim/AgentEnforcer for an example.
By the way, do you know why the project was abandoned? Because when I finished it, it suddenly turned out that the main Gemini 2.5 Pro at that time, like several other models in Cursor, did not support launching with MCP tools enabled! In AI-driven IDE!
The button was returned back. Thanks 
We read and try to fix issues - working on fixing bugs most of this week - some others sent on this thread were also fixed
Thank you for listening, I hope for a good outcome.
Hi @andrewh - do you have any ticket codes you can share with us for some of the bugs we’ve been reporting recently?
Many of us are waiting for things to get fixed, but…
a) we’re not sure if our bug reports have even made it to engineering at all, and
b) we don’t know what references to give you so you can look up progress and let us know if/when these bugs are going to be sorted out.
In my case, the issues with the hooks not doing what your spec says they say they should, and multiple mcp server problems (wrong server calls, and hundreds of stale connections piling up being the main ones, and tls cert problems being a minor one) are my big show-stoppers.