so far so good.. expensive af but still cheaper and faster than Opus 4.6
Thanks Colin what you said makes a lot of sense.
Hi Colin,
How can I gain access to this view, to see the “Auto + Composer” usage available separate from my API usage? This is not present in my “Spending” tab or “Usage” tab. I’m on Ultra plan, exhaust my $400 allowance every month, and had no idea there was a separate bucket of usage I could tap into to try to keep my spending down.
@Colin /others, following up with my impressions after a day of use (105M tokens), most noticeable first.
- Context Window / Management
- It handles long context conversations much better. In particular:
- The compaction process is seamless. With Composer 1 compaction was slow
- It seems to suffer from much less loss of accuracy as the context window grows.
- With Composer 1 I got into the pattern of aggressively opening new chats to optimize accuracy. A prompt that would fail in a long conversation would often succeed if it was the first message in a new conversation. With Composer 1.5, I can stay in the flow and allow Composer to handle context management. I’ve yet to have to start a new conversation to get a successful result.
- Code Quality
- I get less invalid code. Composer 1 was already very good at this, rarely producing invalid code, and when it did, it was minor oversights. But this is even less of a problem with 1.5.
- Caveat: my tasks today did not involve tricky front-end changes/tweaks, which is an area I found Composer 1 to be the weakest on for my particular use case.
- Context: Next.js / React / TS / Tailwind
- I am very curious to see how it performs building, executing, and repairing end-to-end tests. This has been almost impossible for me to do efficiently using an AI in the past, even with Anthropic’s models (which are so damn slow/expensive). If Composer 1.5 can perform well on this, it would be a game changer.
- Speed
- It definitely took its time thinking through problems when necessary.
- There were one or two occurrences where I actually had to check if it was still working because the thinking process was so long, something I’m not accustomed to working with Composer 1 in the past. The thinking time was was well worth it, as it nailed the tasks on the first attempt.
- Overall, I think the speed is still excellent. There were times of the day where it seemed like there was maybe some slowdown due to connectivity or queuing, but that happened with Composer 1 too, nothing concerning.
- Cost / Allowance Consumption
- Cost for a 105M tokens would have been ~$18, but because it’s included in the “Auto” bucket, it only consumed 2-3% of that allowance for the ultra plan (no actual charge). Granted it’s currently discounted, but even with heavy usage, I can’t see myself exceeding the Ultra plan’s auto/API allowances using Composer 1.5. There have been a number of months where I exceeded the $400 API usage using almost exclusively Composer 1 so this is a very welcome change.
- Subagent Usage
- I was expecting Composer 1.5 to use subagents more aggressively, Composer 1 almost never did in my case. While I do not have any custom subagents configured, I find that some models like Codex rely on them heavily. Doesn’t feel like anything is missing or underperforming as a result, just something that was of note.
Hope that helps the team/others looking for an opinion on the new model.
Most of the time it’s unavailable, frequently getting stuck in a script editing state. I have to pause and tell it to “continue” before proceeding, which wastes a lot of tokens.
I used Composer 1 the last month a lot and loved the fast edits in combination with plan mode for quick iteration. It would be nice to have an even faster model for the daily grunt work, and a slower, but more intelligent one, for the difficult design decisions/plans (1.5?). Please keep both options…
Yesterday I focused on GPT-5.3 Codex.
Today I’m using Composer 1.5 — and I really love it for:
-
Speed
-
Intelligence
Honestly, I don’t notice much difference vs Codex for completion tasks.
But I do feel everything finishes faster.
Best part? The visual + multimodal ability. It can see images — that’s a huge help for frontend work. From my perspective, it feels more accurate.
Love how the Cursor team is building this.
I have used Composer 1 pretty much every time since it’s release. Now I am confused with pricing and this auto + composer part as it is not documented clearly.
I tried composer 1.5 on a new feature, it had me pulling my hair out at the end, completely ignored the good practices and existing patterns in my app, I had Shadcn everywhere and somehow decided not to use it, not to use Tanstack form, nor the reusable fields from it, and tried to use native selects 
It took about $17.56 USD to reach a decent result, at 50% discount mind you.
Meanwhile GPT 5.3 Codex took less of my sanity and effort, and ended up taking $14.9 AT FULL PRICE mind you.
So for the time being 5.3 for reasoning for me, Composer 1 when doing simple tasks.
Quality wise It’s actually pretty good, on par or better than Gemini 3 Pro.
I find this to be really confusing. How is this calculated and why isn’t it a specific amount?
This part is actually working for me. My billing cycle just reset yesterday, and after using only Composer 1.5 all day, I see only that bucket ticking up. No idea why yours isn’t 
I’ve been trying out Composer 1.5 all day, so far pretty impressed. Much faster than GPT 5.x Codex, and (so far) the quality has been on par. Composer 1.5 output tends to be more concise and efficient than other LLMs, which I like (less reading for me, and less token costs).
I’d been avoiding using Composer 1.5 because like many I saw the price tag and said “but… why?” Why would I pay more than I would for flagship LLMs that have benchmarks to prove their value?
But after learning about this “auto + composer” bucket thing, that’s a significant value proposition for Composer models. Cursor really needs to advertise this better, because until today I had no idea that was the case (and avoided using Composer 1 for the same reason).
The deal seems to work as promised for me if anyones worried (or at least thats my own experience)
API usage has been at 15% for the past few days, while Auto has been trickling upwards from 1-5%. When I was using Composer-1 I could easily blow through 10% of an ultra plan a day.
Really happy with Composer-1.5 thanks guys!
My experience has been quite bad actually. It actually made my code break, introduced issues through poorly thought out patches (which I didnt ask it to add), which it added instead of actually understanding and fixing the logic.
And I agree with everyone else on the pricing. It seems like the “API usage” bucket just gets filled up super quickly so you are sort of forced to use composer models if you dont want to spend more $$$. Perhaps a tad bit deceptive but not a dealbreaker for me so far.
What is a dealbreaker is how painful 1.5 made my development experience. I cannot depend on it and have moved back to composer 1.0 which actually is better (at least it doesnt introduce overengineered issues)
Thank you all for the great feedback. We do see a large adoption of Composer 1.5
@xiaohang_hu @luchillo17 @asharma would you be willing to share more details?
I suggest using Plan mode when preparing complex tasks as this helps any model.
We do see strong improvements from Composer 1 to Composer 1.5.
Here are comparisons between models and our post on increase of provided usage:
Hi does the “Auto + Composer” bucket include both Composer 1 and Composer 1.5, or just Composer 1.5, it’s currently unclear in all your communications.
I used plan mode for sure, I’m not saying there’s no improvement, but that it doesn’t justify 2x price.
Hello Peter, where are you seeing this nice dashboard?
What I see in my cursor dashboard on web is very different… what am I doing wrong 
I loved the old composer. It was super fast for small front-end changes. Now the new model takes for ever to startup. Even Gemini 3 fast is much faster. Its still dumb so why use this?
Yesterday it was twice faster though.. maybe it’s a capacity issue. Hope they’ll manage to add more resources to handle high loads. Then we should consider that this new model is a thinking one, so it could never be as fast as a non-thinking model like the old composer was.
If Cursor team is listening, please consider to add a non-thinking 1.5 model for simple everyday tasks, that will be as fast as the old 1 model.