1 million tokens does not cost more than 1k tokens for 4.1, as they said in their announcement video, so I have a hard time believing we don’t have 1 million tokens at the moment. Also, 4.1 is currently free in the API, so I would be extremely surprised if they nerf the model that way.
2 Likes
How can you be sure of that? Seems like a hasty assumption from your end.
they have underlying models they are using, simultaneously that are occurring, if they do allow them 1 million tokens, it could easily lead to a ton of requests, backdrop and such, they allow gemini pro max at 1 million context so this has to be intentional
It depends on when OpenAI starts charging for it in their API, and if I remember correctly, they said it would be free for a week.
1 Like
To paraphrase a classic joke from the USSR:
Comrade, good news and bad news: Good news is tokens are unlimited. Bad news is limit is 128k.
5 Likes
Do we know when we will get 4.1-mini to replace 4o-mini?
请问这个问题怎么解决的
He’s not assuming, read at the bottom of this page: Cursor – Models
2 Likes
不知道呀 目前还是处于联系管理员的状态，可能已经是黑号了
1 Like
I see. That is indeed weird since 1m tokens cost the same as 1k tokens, as they said in the live stream. I hope the Cursor team removes this limitation for 4.1 then. I remember reading somewhere that all the requests are being routed through their servers, so maybe they want to keep the traffic as low as possible.
You pay per token, and in Cursor you pay per request, which is 20/500 = $0.04
Which means, if you feed 1 million tokens to 4.1 through cursor, they will have to pay $2 for the input, and even more for the output, let’s say adding up to $4 total for just one call.
As you can see, that is not profitable at all for Cursor and they’d be losing money at a very fast rate, which is why they have a lower limit. That is my take on it.
Funny that if you would use open source competitor tools you could actually take advantage of the 1 Mio Context window via API, also GPT 4.1 seems to be able to find the needle in the haystack with 100% abilitiy which reduces hallucinations drastically and makes large context window use efficient. Hallucinations are by far the biggest problem when vibe coding. So it is just soo bad that they give us here 128k window.
Yes. But the model is completely free to use in OpenAI’s API right now, so not sure why they have to limit it if it’s not for traffic to their servers?
In this specific case yes of course they could use the 1M context window.
There’s several reasons as to why they might not be doing it tho, maybe they don’t want us to get used to the 1M context then they’ll reduce it to 128k once the free-use is over. That’d make us rage even more would it not?
From their perspective it makes sense, they’re a business, they need to profit. It’ll hurt them if they give us something then remove it from us later on, we’ll complain about the 1 million context being gone.
There’s only one thing that I personally don’t understand, and that is: If we provide our own API key, we should be able to use the full 1 Mil context - as the payment for the tokens is then on us.
3 Likes
True. Who knows. Yes, providing our own key should not put a brick wall in front of the requests, and they shouldn’t have to care if users want to spend a lot more money that way.
As I said in another post, but I think Cursor are losing a lot of money currently. $20/month is nothing. I spent about two hours using Cline the other day and I ended up exhausting all of my $20 credits on OpenRouter in that short while. I was just banging my head through the wall and wasn’t thinking about the context length. The Cursor team must take all of this into consideration to both save money and provide a good service for their users. Can’t be easy.
3 Likes
- Very impressive model so far.
- Very coherent comparing to Claude 3.7
- Very good at implementing OpenAI API related tasks
- Not as subservient as Grok-3-beta
They revealed that Quasar-Alpha was 4.1 in the YT presentation.
4 Likes
Same experience, works surprisingly good in 0.45.14 (no way I’m upgrading to latest versions 
)
1 Like
yes. I get that, but technology advances and we are stuck with 128k context window. I understand that they need to make money, but the solution can’t be to limit the user experience with artificial lower context windows (4.1 case related) I’ll will move on from cursor if they don’t handle this problem.
2 Likes