How much is Composer 2.5 subsidized in Cursor? | CodeJam
TLDR
According to my own usage, 10x. If you need more than 10% extra usage<br>from the Pro plan, don’t use on-demand pricing, get a second Pro plan.<br>If you need more than 10% extra of that second one, get Pro+.
In this post:
What I think of Composer 2.5
Composer 2.5 vs. Composer 2.5 (Fast) vs. Auto
What happens when you run out of “Auto + Composer” quota?
How much inference do we get on the $20 Pro plan?
Should I enable on-demand usage or upgrade my subscription?
Conclusion
Context
I’ve been using Composer 2.5 in Cursor since it came out on May 18,<br>2026.
I actually had no clue. I had just ran out of Codex quota and wanted<br>to try Cursor again to see how much output I can get from the “Auto”<br>mode in the subscription.
By pure coincidence, the day I want to subscribe, they release Composer<br>2.5, a significant update to their own model (where “own” means<br>reinforcement learning fine-tuning of Kimi K2.5).
What I think of Composer 2.5
I was instantly impressed by the quality to price feel of that model.<br>It’s one of the cheapest models available, yet it seems to rival<br>GPT-5.5 and Opus 4.8 when it comes to coding? Although it’s tricky to<br>really evaluate that because Composer 2.5 is only available through<br>Cursor, so it’s not possible to rank it on traditional benchmarks like<br>Arena AI’s Code Arena.
Instead we have to rely on Cursor’s own CursorBench.<br>To be fair their scoring of other models is not out of ordinary compared<br>to other coding benchmarks, and the fact the benchmark is private means<br>that it can’t be “gamed” by optimizing for it (except by Cursor<br>themselves 🙃). So if we can’t necessarily trust it for Composer’s<br>performance (it could be trained to perform well on that benchmark<br>specifically without that translating in real-world performance), it<br>might actually be a really good benchmark to compare other models.
That being said in day to day use in real projects, not just toy demos<br>or throwaway stuff, I’ve been pretty happy. The quality is here and I<br>get a fuckton more usage than I was getting with Codex for a similar<br>price.
And that’s despite, with Codex, me trying to aggressively limit the<br>spend by using low thinking most of the time, and older models like<br>GPT-5.3-Codex. Which also resulted in poorer quality and more<br>frustration.
In comparison with Composer 2.5, I stick it to the normal mode (not the<br>“fast” mode they select by default), which is already insanely fast<br>compared to anything else I’m used to (Codex and OpenCode Go), and I<br>don’t have to think about what model or level of thinking to select for<br>every prompt, and deal with the resulting variable quality output. I get<br>consistent quality output and more usage than the competition at the<br>same price point.
Composer 2.5 vs. Composer 2.5 (Fast) vs. Auto
Although there’s no way to know for sure, it feels like picking “Auto”<br>in Cursor often results in using Composer 2.5. Unclear whether it’s in<br>fast mode or not though.
When it comes to pricing, here’s what they charge<br>(price per 1M tokens):
Model<br>Input<br>Cache Read<br>Output
Composer 2.5<br>$0.5<br>$0.2<br>$2.5
Composer 2.5 (Fast)<br>$3<br>$0.5<br>$15
Auto<br>$1.25<br>$0.25<br>$6
So Composer 2.5 (Fast) is 6 times more expensive than Composer 2.5 (it’s<br>not clear to me that it’s 6 times faster though, as far as I’m concerned<br>if I’m not in a rush I’m happy with the normal mode).
As for Auto, it’s 2.5 times more expensive than Composer 2.5, and it<br>seems that you get Composer 2.5 at the end anyway. Maybe a bit faster<br>but no guarantee to be Composer 2.5 (Fast) speed all the time either.
When it comes to the subscription, there’s no indication that the usage<br>you get is proportional to the public pricing of those models. But it<br>would seem logical to me that using a cheaper option results in less<br>quota usage than using a more expensive option.
So as far as I’m concerned since I want to maximize my quota usage, I’ve<br>been using Composer 2.5 explicitly (not through Auto, and not the fast<br>mode) most of the time.
What happens when you run out of “Auto + Composer” quota?
When you max out your “Auto + Composer” quota, if you still have API<br>quota left, Cursor starts to dip into this for Auto and Composer<br>requests.
On the $20 plan, Cursor says that the API quota includes “at least $20<br>of API usage”, so you can expect that from this pool once you start<br>using it. In my experience it’s indeed a bit more, see below.
How much inference do we get on the $20 Pro plan?
Your mileage may vary, but here’s what it came out for me after one<br>month of using (mostly) Composer 2.5.
Auto +...