I saw a tweet claiming Cursor’s new AI model, Composer 2.5 released on May 18th, matches Claude Opus 4.7 at a tenth of the price. My first reaction was…WORD?!! Then immediately: I have to test this myself.
I am on Claude Pro and I hardly ever touch Opus 4.7. It burns through my session limit before I have finished a thought. So when I saw that claim, I stopped scrolling, upgraded to Cursor Pro, and ran a real test. Not a hello world. The exact same complex task I had already built with GPT-5.5, so I had a direct comparison waiting.
Here is what happened.
What Is Composer 2.5
Composer 2.5 is now available in Cursor. It is a substantial improvement in intelligence and behavior over Composer 2, better at sustained work on long-running tasks, follows complex instructions more reliably, and is more pleasant to collaborate with.
The numbers behind that claim are worth looking at. The cost is $0.50 per million input tokens and $2.50 per million output at standard tier. Claude Opus 4.7 costs $5 input and $25 output. That is exactly 10x cheaper. Not marketing rounding. Just the maths.
Composer 2.5 is built on the same open-source checkpoint as Composer 2, Moonshot’s Kimi K2.5. Together with SpaceXAI, Cursor is training a significantly larger model from scratch using 10x more total compute with Colossus 2’s million H100-equivalents. That future model has no release date. Composer 2.5 is what ships today.
What Is Actually Under the Hood
Composer 2.5 is not built from scratch and that is worth understanding. The base is Kimi K2.5, an open-source model from Beijing-based Moonshot AI. When Cursor launched Composer 2 in March, the community discovered the Kimi foundation by spotting the model ID in API response headers. Cursor had not disclosed it clearly at launch and the community reacted. This time they named it upfront.
Cursor confirmed that roughly 25% of the model’s computational foundation derives from the original Kimi K2.5 architecture. The remaining 75% is Cursor’s own work: continued pretraining on coding-specific data, then large-scale reinforcement learning on top of that stronger base.
Composer 2.5 was trained with targeted textual feedback, providing feedback directly at the point in a trajectory where the model could have behaved better, targeting specific mistakes like invalid tool calls or poor coding behavior. That is meaningfully different from just fine-tuning a base model and calling it done.
Think of it this way. Kimi K2.5 is a capable generalist. Composer 2.5 is that foundation after intensive specialist training in one domain. For coding tasks specifically it outperforms the base. For general reasoning and long-context work, Claude Opus 4.7 still leads. The benchmarks are coding-specific. That context matters.
How You Access and Pay for It
Before the test, a practical note because this tripped me up.
Composer 2.5 requires Cursor Pro at $20 per month to use properly. The free Hobby tier gives you limited Agent requests and Tab completions, enough to evaluate the tool but not for real work. On Pro you get extended Agent limits and access to frontier models including Composer 2.5. Additional usage beyond your included amount is billed on demand. For most everyday coding work the included limits are generous.
One thing I noticed during the session is that the context window fills up and resets as the agent works through long tasks. Cursor uses a context compression approach so the model keeps working rather than stopping when it hits the limit. For some developers used to Claude’s much larger context window this feels like a ceiling. For me, coming from a Claude Pro plan where I ration Opus sessions anyway, it was manageable. The model kept its thread through multiple resets without losing the plot noticeably.
Some developers find the 200k ceiling limiting, especially those used to Claude’s 1 million token context window. For me, coming from Claude Pro where I avoid Opus because it burns through my session limit before I have finished a thought, the 200k reset was not a dealbreaker. The model kept its thread through the resets without losing the plot noticeably.
The Test: Same Prompt, Different Model
I wanted a real comparison, not a benchmark number. So I took the exact prompt I used to build an interactive cloth physics simulation with GPT-5.5 and ran it through Composer 2.5 unchanged.
The task was building a complete Interactive Cloth Lab as a single HTML file. A flag with August Wheel branding, billowing in a cinematic dusk sky, with animated birds, working physics controls for wind, turbulence and stiffness, all rendered in the browser with no dependencies. The GPT-5.5 version, live at cloth-lab.vercel.app, used raw WebGL and produced a polished result in two iterations. That was my benchmark.
The first thing I noticed with Composer 2.5 was the speed. It moves fast. The Thinking panel inside Cursor showed the model planning its architecture before writing a single line, reasoning through physics constraints, rendering approach, and UV mapping. The intelligence was clearly present.
The first result needed work. Several iterations followed, each one meaningfully better than the last. I pushed it on tool choice, rendering approach, and physics constraint specifics. Each time it corrected correctly without introducing new breakage elsewhere. By the end, the flag was billowing convincingly against a dramatic generated sky background, the physics were holding, and the scene was genuinely shareable.
It was not a one-prompt result. It took guidance and iteration. But the speed of each iteration and the quality of reasoning at each step made the process feel collaborative rather than frustrating.
The honest comparison: GPT-5.5 reached a polished result faster on this specific creative rendering task. Composer 2.5 got there with more direction. On raw coding tasks closer to its benchmark conditions, the gap would likely narrow significantly or flip entirely.
What This Means for Everyday Builders
The AI coding tools market is moving in one direction: specialist models trained on top of powerful open-source foundations, competing with closed frontier models at a fraction of the cost. Composer 2.5 is the clearest example of that trend yet.
If you have been avoiding Claude Opus because the cost burns through your budget, or you are on a $20 subscription and want more from it, Composer 2.5 deserves a proper look. Not as a replacement for everything, but as a dedicated coding environment where the cost to performance ratio is genuinely compelling.
Many developers use Cursor for everyday IDE editing and switch to Claude Code for complex autonomous tasks. The two tools complement each other more than they compete.
The future is here. It is just not evenly distributed yet. But at $20 a month, it is getting closer.
Frequently Asked Questions
Is Cursor Composer 2.5 better than Claude Opus 4.7? On coding-specific benchmarks Composer 2.5 matches Claude Opus 4.7 closely and costs 10x less. For general reasoning, writing, and long-context work beyond coding, Claude Opus 4.7 still leads. For coding tasks specifically, Composer 2.5 is the better value by a significant margin.
How much does Cursor Composer 2.5 cost? Via API, Composer 2.5 costs $0.50 per million input tokens and $2.50 per million output tokens at standard tier. On Cursor Pro at $20 per month.
Can I use Composer 2.5 on Cursor’s free plan? Cursor’s free Hobby tier includes limited Agent requests and Tab completions but does not specify exact numbers publicly. In practice the limits are tight enough that meaningful testing of Composer 2.5 requires upgrading to Pro. I found this out myself before running the test in this article.
What is the context window for Composer 2.5? 200,000 tokens, equivalent to roughly 15,000 lines of code. When the context fills, the model compresses its own context and continues working.
Leave a Reply