Composer 2.5 vs Opus | The Results Are Brutal

追加: 2026-05-20

[00:00:00]Introducing Composer 2.5 from Casa on par with Opus 4.7. That is a huge jump.

[00:00:07]Even Elon Musk tweeting about this because this is strained on Colossus 2, world's biggest supercomputer with 200,000 GPUs. And this Composer 2.5 is trained on that. Also, the cost is comparatively cheaper than other frontier models. We are going to see in detail and also we are going to try it out. And I'm going to show you how you can test it yourself and see how it turns out to be. That's exactly what we're going to see today. Let's get started.

[00:00:39]Composer 2.5 in terminal bench is 69.3.

[00:00:44]Opus 4.7 69.4. SWE bench multilingual is 79.8. Opus 4.7 80.5. This is very close.

[00:00:54]Casa bench 63.2. Opus 4.7 is 64.8. So, you can see it's competing with the top-performing model. If you see the cost per task comparing with Opus 4.7 and GPT 5.5 Composer 2.5 costs very little considering the task it's trying to perform is nearly equal to Opus 4.7 and GPT 5.5. Also, this is trained on the open-source checkpoint of Moonshot Kimi K2.5. Kimi K2 is here. K2.5 is here. And then Composer training and reinforcement learning is here. The way they improved this is by textual feedback inserts, targeted hints to learn corrections. Agent rollout with hint. A reminder is available tools are read, write, shell, and string replace.

[00:01:46]The hints are provided like this. If token possibilities that violate the hint go down and the model weights are updated to avoid this error. Next, they use synthetic data to train this model.

[00:01:57]25 times more synthetic task than Compositor 2. I've been using Compositor 2 for a long time now, and I really like it. And considering we got Compositor 2.5 on par with the top model, it's going to be my go-to. Synthetic data is used to create harder tasks to increase model ability. Compositor 2.5 is priced at $0.5 per million input token and $2.5 per million output token. That's comparatively cheaper. And if you use the faster variant, that is $3 per million token and $15 per million output token. And it includes double usage for the first week. To use the model, go to cursor.com, download it, and install it.

[00:02:40]There, when you go to the chat interface, at the bottom, you got Compositor 2.5. The faster version, if you want to switch back to the slow version, you can switch it like this. By doing that, the cost is going to be $0.5 per million token. But let's try with the fast version. I need to do some security audit for one of my application. So, I'm going to say with the plan mode, do security audit and fix and create pull request. And then clicking enter. This is one of my application, Prazen AI. So, it's going to go through the issues. It's reading faster. It prepared me a plan, as you can see here. That was very quick, just within 30 seconds. And now I'm going to build. Clicking the build icon. Based on this plan, it's going to fix all the issues that it identified. You can even see the progress in the plan area. That it's loading now. It's going through the first one. And now it's making those changes. It's running some tests to make sure that everything's working as expected. As expected, it completed all the task and finally created me a pull request after fixing all the issues. Do try and let me know in the comments below what do about this. Considering you already like Corsair, I've also created another video. It's about Corsair's other features. I'll put the link in here and I highly recommend for you to watch and I will see you there.