Formula 1, Cancun, Mexico, 1 August 2015. View public domain image source here

Cursor’s new Composer 2.5 model is fast, cheap and amazing

Cursor released a new model called Composer 2.5 and turns out you can get something fast, cheap and with a high coding quality. We really are spoiled.

This comes hot off the heels of the announcement that SpaceX has an option to acquire Cursor for $60 billion.

Not a Medium member? Keep reading by clicking **here**.

Elon gives an indication that acquiring Cursor is a wider part of making Grok better.

On the Cursor bench, it’s interesting as their average cost per task is extremely low at just $0.55.

That is $0.55 compared to $11.02 for Opus 4.7, a 20x cost decrease and $3.59 for ChatGPT 5.5 High — a 7x cost decrease.

On ChatGPT’s own self reported benchmark, they claim it is the best. So we have to take that with a grain of salt, as it seems everyone these days says they have the best models.

So I have been testing the model and seeing how it compares as a normal person.

First up, I have used 209 million tokens in the last 4 days.

I have been using the tools quite a bit and it only used 4% of my allowance for Composer and 13% of my API allowance for using models like ChatGPT 5.5 and Opus 4.6/7.

So this answers the first question, this is extremely good and valuable on a cost per task basis and that $0.55 metric does feel to me to be fair that Cursor has cited.

I recently wrote about how the $20 and $60 plans are a joke.

[embed]Claude Code & Cursor — the $20 and $60 plans are a joke I soon fear the $200 a month plan will be a joke too. We are getting rug pulledmedium.com

Now that we have Composer 2.5 I think that at least the $60 plan is usable. I tested the $20 plan on a friends account and still zipped through my usage allowance so quickly, so I still think $20 plans are a joke. Maybe I don’t really understand hobby coding anymore as I’ve got used to just coding big projects and maybe the market for the $20 plan is just a dabbler.

Let’s talk about speed

Let me be clear, Cursor 2.5 is the fastest model by a country mile. It is like a rocket ship.

It is so fast that it is honestly worth switching to Cursor for.

The speed is crazy and it also performs at a high level. As soon as you write a prompt, it starts to run multiple things in parallel and so many of them just work, it’s incredible.

I wish every AI product was this fast.

The speed makes you code more things because it gives you more ideas and helps to keep you in a flow state.

The difference between this model compared to the previous Composer 2 model

I used the previous Composer 2 model heaps and I used it as a task execution model. I basically would just use Opus 4.6 for planning and then I would use Composer 2 to build the plan.

Now I just use Composer 2.5 for planning and for building. It really is that good. That makes a huge difference as now it brings your coding bill dramatically down and also ups how much you can get done.

The difference here is that you can use it to plan and think. I just could never get Composer 2 to do that, it would always do strange things and make very strange UIs and just be all over the place.

I honestly haven’t noticed the difference using this compared to Opus 4.6 for planning.

The main test for me has been that the features work.

So what are the implications of this?

Let’s look at this from a few lenses

If you are into Claude Code but don’t love using the terminal

Then you can definitely swap to Cursor, this new plan solves the big problem of having a model that is affordable that you can run for ages and it is really fast.

If you are deep into Claude Code and love the terminal

It depends here, do you love speed? If so, add the $60 plan and test it out for one month. If you don’t love speed, then you could stick with Claude Code.

What about if you use Codex?

I think Codex plus Cursor is the best combination right now and that is what I use. I enjoy the UI of the websites and look and feel that Codex makes and it feels very easy to just code and get lost in the flow.

But that speed is addictive from Cursor.

So I now have a $200 a month plan on Cursor and a $100 a month plan on Codex.

I wonder if Anthropic will follow suit

Anthropic will be paying close attention to this and if they release a fast model, we will have to reassess everything again, I am sure that they will in time.

Anthropic is probably working hard to make 4.8 better than 4.7 as everyone hates 4.7 and it’s not good for their brand to have a model that people hate. Maybe they will just cut the cord and jump straight to Opus 5 and make it a clean slate.

OpenAI is in a good place with ChatGPT and on Codex you have the fast mode. The only thing with that though is that it chews up the tokens much faster than Composer 2.5 and I have found that ChatGPT 5.5 on medium normal mode is perfectly fine for most tasks.

I’m not affiliated in any way to any of these companies. I write these as a normal person helping you to navigate what stuff is good or bad in plain English without going over the top.

Before you go

I’m looking for beta testers for my book recommendation site. You take a quiz and then it gives you books tailored to your preferences.

https://www.readerprint.com/

If you find that Goodreads and Amazon don’t give you good recommendations then this is for you.

Also feel free to sign up to my Substack. I do a unique post there every week and I also post a link to all my Medium posts every week.