The AI coding race is heating up again. After OpenAI, Anthropic, and Google, Cursor has stepped into the game with its new model, Composer 1, a coding-focused agent model that’s said to be 4x faster than other models with similar intelligence. 🤨

It’s said to output code at lightning speed, reason through large contexts, and even outperform models like GPT-5 and Claude Sonnet in engineering workflows.

That’s a bold claim, so I decided to test it myself. In this post, we’ll see how Composer 1 performs when building an actual agent and to make things fair, I’ll put it head-to-head with Claude Sonnet 4.5, one of the most consistent coding models out there.

TL;DR

If you just want the results, here’s a quick rundown of how both models performed in building a simple agent:

• Composer 1 produced the most complete implementation and had the fastest output. It coded the entire agent in under 3 minutes, though it needed two small follow-ups.
• Claude Sonnet 4.5 also got the job done, but sometimes used outdated API methods, even though I clearly provided the latest documentation for a Python package. It tends to rely more on its training data than the instructions you give it.

In terms of code quality and implementation, there’s not much difference. That said, Sonnet 4.5 burned almost twice the tokens, while Composer 1 delivered similar results in half the time (not 4x) with far fewer tokens. It’s efficient, fast, and feels like a strong pick for everyday coding.

Brief on Cursor Composer

Since this model dropped only a few weeks ago, here’s a short refresher.

Composer 1 is the first agent-focused coding model from Cursor. They claim it’s about 4x faster than similarly intelligent models.

(By the way, what even is the Cursor Bench Score? Can't really trust all the metrics blindly. 🤷‍♂️)

It is said to be a mixture-of-experts (MoE) language model supporting long context generation and understanding. As mentioned earlier, it's built especially for building agents or general software engineering workflows in mind through Reinforcement Learning (RL).

They've positioned it as a frontier coding model, with top-tier intelligence that seems to surpass all its peer models with similar intelligence. It comes with a 250 tokens per second output speed, which is roughly twice as fast as most coding models and about 4–5 times faster than some reasoning models.