Optimizing code bases with proper code splitting and rules for models to be more efficient will be the new lucrative position, im assuming you could also write rules so that the choosing of model is dependent on the type of task, sort of like auto but tuned for your preferences
You needed to have bought an annual subscription before they sunsetted the old plan.
thank u for the tips but we are startup, so i am super active user, like 12-16hours coding always, mulitple chats) that’s why
My process is very efficient relative to the high token usage some are reporting. Yet, LLMs do 99% of my coding on production systems. The key is I have rules guiding them to efficiently parse hierarchical documentation. For large feature sets, we create a requirements doc, then create a detailed implementation plan from that before they write a single line of code. When I complete a new feature, we review and update the documentation.
While this has allowed my projects to succeed, and I have not hit the high edge of token use, it has not prevented me from being impacted by the price squeeze like everyone else. In other words, I don’t think there is anything we can realistically do to return to previous productivity levels at the same cost today, though I’m hopeful it will improve in the future.
If we spend lots of time manually building context and writing precise prompts just to try to lower token cost, that eats away at our productivity.
The only thing we care about is how much it costs to reach a goal. While we can change methodology to try to keep cost from skyrocketing, until drops in inference costs are passed on to us, productivity will not soar as many of us cannot just accept rapidly rising costs.
Yes, you pay for api use. The key difference is that you have immediate and easy access to all proprietary and open source models via Kilo Code’s api router. By doing most of the coding with the cheapest most-advanced models (Grok Code Fast 1 and GLM 4.6 at the time of writing), I barely spend like 2 dollars in a day of intense work. The crux comes when I have to rely on Claude Sonnet 4.5… THEN you can see your balance shrinking in real time. One possible solution to this is to have a basic subscription to Claude Code so you can use this advanced model when problems arise. It’d be a small investment for a year, since cheaper models would eventually catch up.
I just couldn’t trust them to honor that commitment to do that. Instead, I have dramatically reduced my dependence on Cursor by using codex, engaging ChatGPT more and completing non-coding objectives while I continue to test non-Cursor opportunities.
I agree, i wasn’t really countering your point just wanted to say that codebase optimization for LLM purposes will be a big thing.
holy moly how did you do that
Weird. Last time I used Trae IDE it was just yet another garbage VS Code clone wrapped in a fancy theme. I did not have the agent features so I uninstalled it. Also privacy of those random no name tools might be questionable. At least cursor is very transparent. They state openly that they have pieces of my code in their embeddings in the DB, they don’t store the code in legacy privacy mode. That’s def RESPECT+ as I hate when software changes privacy related stuff without letting me know because it’s convinient for them to basically steal while I’m not watching. Cursor doesn’t do it, what about Trae IDE?
Are you unaware or ignoring the fact that the models have started generating more tokens to complete new tasks?
This, so much.
It’s crazy to me how much more tokens are burnt, but I genuinely do not feel that the quality or efficiency of LLM agents have improved even close to warrant it.
Hell, I feel like over the past year the main thing that has happened is that stuff goes slower, costs more and produces more unstable results.
Like, I don’t feel that anything has gotten significantly better since Claude 3.7 running in Cursor 0.46.
So many new versions, but it feels like I haven’t gained any significant improvement to the actual tools I work with to improv my workflow.
Is it really insane ? What is the value you get out of the 200$/month subscription even now with no free Auto ? It’s by far more than 200$ in my opinion and as a startup investor I think at a point in time every startup needs to stop burning money and start to earn some (or at least burn less). Even if you get half the tokens next month it’s still worth the money and I guess cursor is still losing money with every sub they sell…
Nevertheless communication wise the new pricing scheme is a mess. And 99/100 startups underestimate what it means to fail on customers (our) expectations.
This. I was fine with the older versions of Cursor where less tokens got the job done. Now they’ve overkilled the app and tell us to manage our token usage better. I canceled my subscription months ago. There’s less overkill apps that get the job done for reasonable money. Cursor unfortunately just isn’t an option, even though i’ve achieved great results with it. What a shame.
I was talking about the evolution of models and the fact that they now solve more complex tasks that require more tokens. It doesn’t depend on how Cursor works.
It’s directly tied to how Cursor works. It allows the LLM to utilize all kind of tools within the app which often isn’t necessary for the task. People on the forum are also complaining about token burnage on .MD files, even when stating in rules not to make one. That’s just one example. There’s just too much extra, while the older versions of Cursor were simpler, communicated simpler with the LLM, but got the job done. They should just add a button imo which makes the entire Cursor app “simple” or “expert” mode - often simple’s enough for most of us.
All the models that I use create documentation only upon my request (if that’s what you about). You can try my Agent Compass instead of your own rules for comparison.
Cursor Team has already broken web_search and terminal. I would not like any additional simplifications.
Also, I can’t say that any functions in Cursor are superfluous. There are some that I don’t use, but I still understand who they’re meant for.
It’s still Cursor burning tokens, not the LLM’s.
Now Claude is approximately 3.5 USD or 30-60 minutes of coding every 5 hours, one session is about 10% of the weekly limit, so 35 USD per week for the 20 USD per month plan. I am using and mixing Cursor and Claude.
I have an account that subscribed to an annual plan 4 months ago, and it’s still at the old 500-limit. If you don’t switch back to the old price within the specified time, you’ll never be able to switch back.
Intentionally to pretend to be stupid and intentionally forget things, and talk you you with a very confirm tone to something that is absolutely wrong, in which such thing been done 50 times before