me too, and I give out, and now back to grok-code-fast-1 now.
Yea, 4.5 is super token hungry and expensive. I think people are learning to use it sparingly and instead use gpt-5 or other models that are more cost effective.
cursor offer most if not all Claude models at the same cost of direct API. The reality is as these models get smarter and more capable they get more expensive. Whilst i agree it is getting quite pricy its absurd the amount you can get done in just a day with 4.5.
I agree. I am far more careful in choosing the best tool for the job now.
I used to use Claude 100% of the time. Now I find Claude is just 10x (or greater) more expensive and is not the smartest in design, testing or even engineering. I find its only value is in reviewing code. GPT-5 (medium or high) is now my go-to for all complex tasks and grok-code-fast-1 for as much as I can since it seems just about free. I actually use it the most time-wise (and see how much that costs at the bottom here).
Here is an example of Claude failing to make a simple utility function that strips some objects out of data. Very simple task. It made a mess and created unittests that made no sense. I switched to gpt-5-codex and it did the task with better design and really good tests! This is the cost. Seriously, the same task:
Another example? I created a summary prompt of about 50 issues from a CodeRabbit review and thought Claude would be the best to nail them all, ranging from simple unused imports to real issues that need engineering thoughtfulness (I told it to never trust and always verify fixes, etc). More often than not, fixes required manual clean-up or re-design anyway, which I continued to use it for. Claude cost me $17! (In a little over an hour)
Now for code review “auto”-fixes, I find it much better and faster and cheaper to let grok-code-fast-1 do the first pass, then let gpt-5-codex review or handle the complex fixes.
In fact, the only thing I’ve seen Claude Sonnet 4.5 do better than GPT-5-codex recently is TypeScript typing, as GPT-5 loves to over-type and use escape hatches (...as unknown as...). I just posted a “Claude Sonnet is dead to me” message in Slack to my team (with the info above). Largely it is the PRICE. I would use it more to see where it is the right tool but it is so expensive I don’t think it ever is now. Especially now that we have codex, which has variable reasoning and can answer quickly when thinking isn’t needed.
Much of this is a new pivot for me…maybe in the last week.
The point is, that single code review I used Claude for doubled my bill at the time(!), versus working smartly with grok-code-fast-1` and `gpt-5-codex`(which really seems to have gotten better in the last few weeks!)
28m tokens with two prompts. Impressive.
Oops, I should have stated that differently. The first $17-ish dollars was a code review where it probably spent about $6 on the first prompt and the rest is me using the model all the way through corrections and redesigns to get to the total. Still, that would have been more like $3 for grok-code-fast + codex usage.
Currently it pretty impossible to use Cursor.
Not that is not good. Its amazing. BUT
- Trying to use a cheap model, such as DeepSeek, tooling does not work as well as with ChatGPT, Gemini. and Claude. So only good for output that is not that big.
- Using any of the big guns, is expensive and eats up credits very fast. It will not last more than a week (and I am generous)
- What are the alternatives? For me, I already have a GPT Plus subscription, so codex can be used there, + gemini cli + paying for claude code (instead of cursor) I think gives me more for the buck
Of course the providers can offer a better price. That’s obvious. For now, I will try that modeland keep my options open). I am not sure which configuration will be best.
Any feedback for those that tried that solution is really welcomed.
i tried out the playwright mcp with 4.5 and managed to use 2m input tokens in 1 prompt. thats the most ive seen it do so far costing ~$6.
yup
sonnet ■■■■■ in terms of pricing, agents nowadays burn tokens like a forest fire so 3/15 per mil is ridiculous
if you wanna use it better to use products with subsidized costs i.e. not fair to the provider to “rug pull” users to more fair pricing later
for cursor I recommend using gpt-5-mini and grok-code-fast actually, they are cheap as hell and get job done with decent enough quality
I never reach limits. partially because I rarely use agents but also cuz I don’t use sonnet at all 
I think we are going to continue seeing the trend that the high token users are gonna start pricing themselves out of these models they became dependent on as the market settles. There are a lot cheaper options like you said, but Sonnet does do a more thorough job. For those that need that more powerful model because they are not programmers are going to be forced to pay a lot more money in the long run.
It was probably the plan all along with these advance models (heavily subsidize, gain market share and dependence, then increase the price now that people have no choice). I am like you, I use agents like 30% of time (mostly tab auto complete). And when I do use an agent, its very specific where I can usually get away with using something cheap like gpt-5-mini or grok-code-fast, and my Cursor usage stays very low.
Cursor is a great deal for someone like me, but then there are people dropping $12 on a single prompt, which is an insanely different use of the tool, and its going to be very very expensive to keep that reliance up. And as we see, people are going to rage quit Cursor but there really aren’t a ton of cheap alternatives for these heavy vibe coders.
This is only possible if you switch from the Max mode of the million-token model to the normal mode. Otherwise, only cached tokens are lost, since the models have a similar context window in normal mode.
It depends on the task. gpt-5-high can study the repository for 10 minutes without changes, but then produce good results. According to the information that ChatGPT found, -low should be about 15% dumber than -high.
Because of the cost, I can’t afford a full test, but I don’t feel like it’s that much smarter than gpt-5 compared to how much more expensive Sonnet is.
I had a long discussion with GPT, and it appears that inference costs, particularly in terms of chip costs, are dropping dramatically every year.
Yet, Anthropic and OpenAI have yet to make a profit. NVidia, on the other hand, is making a killing.
So, two things could happen over time:
- They begin to pass the dropping inference cost on to us.
- Competitors to NVidia allow inference companies to get some of their profits through lower costs.
So, while the Cursor pricing was promotional and then it wasn’t, if they want market share and have cash to burn, it is in their best interest to let time be their friend instead of racing towards profitability. Fortunately, if they aren’t wise, we have a lot of choices popping up every month; so economics will over time favor developers. Prices do not need to rise.
This price is indeed a bit frightening.
Are you unaware or ignoring the fact that the models have started generating more tokens to complete new tasks? Last year, I couldn’t even imagine a task where LLM would need to process 15 million tokens for me, but now it can be a common refactoring task.
holy tokens…
you might wanna look into optimizing your workflow with AI tools, it’s not only Cursor who will “rag pull” you in the future
see Complex Context - TIP! for example
agents don’t operate in “requests” that can be predicted it costs accurately, they loop until the job is done, so they may take a s**t ton of turns burning millions of tokens just for a simple few lines change
it is partly on Cursor as they are not optimizing their meta prompts enough but I believe they did a pretty good job already as I don’t find myself in these situation, so it might also be the way you use it that causes insane spendings
AI is the future, and it’s in your best interests to learn to use them as effective as possible
It would be great if it were true…
Different users use Cursor in different ways. They can’t make the system prompt too smart, because it will cause problems for some users.
true, that’s why we as users should develop our own ways to make it work while Cursor devs should focus on making it as universal as possible, maybe even with options to “customize” some aspects of their meta prompts like ticking on and off certain features in it or selecting different variants of built-in mechanisms