I want to be cost effecient while i don't want to waste my time fighting the keyboard on some tasks the AI can't uderstand, i know claude fable or at least opus would be the fit since they're the best in the market right now, tho, i want to know if for example cursor might be a good start i'm afraid of tokens limitations, so if i'm gojng with claude which plan to choose, i'm also hearing grok is now great for coding
Guys this is the first time i'm going to use AI for agentic work, coding and hading things with MCPs, which model to start with
- Thread starter anasbinti
- Start date
Mendoza952
Registered Member
- Aug 25, 2026
- 51
- 7
honestly the cursor $20 plan is probably your best bet because you get unlimited slow requests anyway so token limits arent really an issue. grok is getting better but sonnet 3.5 is still the king for coding imo.
qwen still nothing compared to claude or am i wrong because it's open source and from what i can see people are not very satisfied with the results, just also wantto mention that if you can recommand any qwen model host provider so i can use it as an API call
do you mean just with 20$ i can keep code with it without limits ? or am i not understanding things right, never used it on the paid plan, tho with the free plan i was very happy with the results, can it also be used for other agentic process same as claude ? never being using these AI agentic stuff so i'm still noob with using all of these tools, for example is it possible to integrate the models from cursor on vs code, add skills/MCP tools ?
Last edited by a moderator:
Nacho Varga
Newbie
- Feb 8, 2025
- 10
- 10
Codex is really good
Nameless Knight
Registered Member
- Feb 6, 2026
- 94
- 41
Use an SLM for agentic automation, something like Llama or Gemma
Its faster, cheaper, and less mistakes as long as the tasks are simple in nature
Its faster, cheaper, and less mistakes as long as the tasks are simple in nature
- Sep 10, 2010
- 11,838
- 26,690
I know claude is pretty much the gold standard, but qwen being a free model isn't as bad as people think. Try using it with a harness e.g. something like Hermes Agent. I actually have used harmes agent with qwen coder. Worked surprisingly well for being a free and local setup.
As for the API provider, I am not really sure. I just use it directly using LM studio on my m1 max macbook.
flackz
Senior Member
- Nov 28, 2023
- 846
- 351
If you are going to use LLMs for these tasks, you will fail. That's an honest and straightforward answer to your question.
couldn't understand you answer, i'm just trying to code with AI what you mean LLM ? we always using LLMs to code with ? we just feed them data in a certain way or am i wrong ?
flackz
Senior Member
- Nov 28, 2023
- 846
- 351
The fact is that there is no such thing as AI. What is commonly referred to as AI in the media is actually large language models (LLMs) based on the Transformer algorithm. These systems cannot think in terms of the physical world, and they have absolutely no understanding of programming. All they can do is continue the given text (prompt) given to them indefinitely until they are forcibly stopped. Think of LLMs as advanced versions of T9, which was previously used in cell phones.
Please read this article on how LLMs generate their responses.
Please read this article on how LLMs generate their responses.
but people already code with it, and i've already built some scripts with it which makes my work a lot easier, what you mean ?
flackz
Senior Member
- Nov 28, 2023
- 846
- 351
If you try this yourself, you'll see what I mean. An LLM is trained on a massive amount of code from the internet and can easily generate a boilerplate application based on your request. It will be exactly that boilerplate code, more or less similar to what you wanted. But if you try to refine that code using the same LLM, your code will fall apart with each iteration. As a result, you'll end up with some sort of Frankenstein's monster that somehow works but contains a huge number of hidden bugs (lurking until a certain event occurs) as well as security vulnerabilities. Maintaining and refining such code will be simply impossible.
I can relly to this even thought right now, agents can solve the code problems, i know "AI" it's not pefect but it can get the job done you won't hvae to call a dev for something doesn't need to much of expertise
xReminisce
Elite Member
- Dec 29, 2012
- 7,344
- 36,683
claude > ChatGPT > Qwen > DeepSeek > Cursor > Gemini > everything before fucking Grok
but ur gonna hit 5hr limit pretty fast which probably would frustate you
cursor doens't have the 5hr limits, but from personal experience, since elon fucktard took over, they have been aggressively shilling their new shit model grok and it's kinda doing whatever tf it wants.
Claude is much better for handsoff agentic work.
with cursor you don't hit the limit or something because i see people say you won't find a limite issue with cursor but you guys mean, they have limited tokens right ?
xReminisce
Elite Member
- Dec 29, 2012
- 7,344
- 36,683
Not true.
At $20, i'm assuming you are getting the cheapest plan right?
Cursor gives you roughly give or take 700million tokens for the month. This includes your 3rd party API usage.
I said roughly because it depends on your model, the more expensive model, the slightly less you give. But its usually around the 700 million range.
Then it gives you roughly 130million tokens on Grok bot per week. So that's an extra ~500million tokens for GrokBot standalone. These 2 pools do not mix.
First month of usage, after you finish your standard ~700million tokens, they give you something called "Slow-Mode" AUTO. Which is just their regular auto but your on the "free queue", same model, same stuff, just it may take more than 10sec before your prompt gets to their servers to be processed. longer if during peak time.
Second month, they reduce or sometimes just outright disable that "Slow-Mode" AUTO. so once you reach 100% usage (~700million tokens) it will tell you to upgrade or u cannot use it.
BEFORE elon took over, that "slow-mode" was always there, even if u subbed for 2 years.
That's probably where you are reading the "unlimited" tokens and no limit
how long does it take to get the 5hrs limit reseted ?
xReminisce
Elite Member
- Dec 29, 2012
- 7,344
- 36,683
5 hour start ticking from the moment you type your first prompt.
It's a rolling 5 hour window.
so if you start at 1PM, it resets 5hr from then.
However, then you have a "Weekly" Usage.
Which means, doesn't matter if your 5hour window isn't 100%. Weekly is the true "Cap".
Iirc, for a $20 plan (sorry, not on a $20 plan for claude anymore), 100%ing a 5hr window is roughly 5% of your weekly. (or was it 10%... i cannot remember completely sorry)
Which may increase by roughly 25% due to their promotion ending soon.
Since i'm just scripting things and asking for help, maybe the most advance thing i will do is using MCP server for server diagnose or something like this, do you think claude 20$ offer would be enough as a start ? really just trying to get the model to read some documentations and build on top of that, linking APIs and so on
xReminisce
Elite Member
- Dec 29, 2012
- 7,344
- 36,683
Yes, claude should be enough
Claude does have the 5hr limit.
However, it does give better models overall imo.
Token wise, including caching. You are getting ~4billion a month compared to cursors ~700m+500m
so alot more value for money.
Just you cannot do mega long stretches of work without it telling you to touch grass or pay API pricing.