Cursor Composer 2 cuts coding-model prices, but the Kimi base changes the story
The hero scene emphasizes Composer 2 as both a pricing attack and a story about an initially hidden model base.📷 AI-generated / Tech&Space
- ★Cursor reports 61.3 on CursorBench and 73.7 on SWE-bench Multilingual for Composer 2
- ★Standard pricing is $0.50 input and $2.50 output per million tokens
- ★Cursor later acknowledged that Composer 2 builds on Moonshot AI's Kimi K2.5
Cursor's Composer 2 is not interesting because it claims to be the best coding model. It is interesting because the standard version costs $0.50 per million input tokens and $2.50 per million output tokens. The faster variant, which Cursor describes as the default, costs $1.50 and $7.50 per million tokens.
That directly pressures the economics OpenAI and Anthropic have built around premium coding models. Cursor's own announcement lists 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0, and 73.7 on SWE-bench Multilingual. The numbers are useful, but they should be read as Cursor's metrics, not an independent verdict.
The Decoder added the key turn: Composer 2 was not clearly presented as a model built on Moonshot AI's Kimi K2.5. Cursor employee Lee Robinson later said roughly a quarter of the pretraining came from the base, while the rest came from Cursor's continued training and fine-tuning.
That does not disqualify the model. Fine-tuning a strong open base for a narrow domain is often a smarter path than trying to build a frontier lab from scratch. The problem is that the market hears "our model" differently when it learns later that someone else's foundation sits underneath.
Composer 2 looks like Cursor's answer to OpenAI and Anthropic, but the important lesson is not the benchmark; it is the value of transparent fine-tuning on someone else's open model.
The second visual shows the distinction between an open base, fine-tuning, and the market product.📷 AI-generated / Tech&Space
Composer 2 targets long coding tasks where an agent has to perform hundreds of steps, not just complete one function. Cursor therefore emphasizes reinforcement learning on long-horizon coding tasks. That is relevant because modern IDE agents increasingly resemble small operators that touch files, terminals, and tests, not autocomplete boxes.
CursorBench, however, is not your legacy monorepo. Synthetic and internal benchmarks can reward tasks that are tidy, isolated, and well-instrumented. Real production code has half-documented APIs, messy migrations, security exceptions, and human requirements that arrive in the last sentence of a ticket.
The strongest argument for Composer 2 may therefore be distribution rather than raw capability. Cursor already sits in the developer workflow, supports third-party models, and can decide when its own model is good enough. If a cheaper model is good enough for a large share of tasks, premium models become specialist tools rather than the default.
The conclusion is uncomfortable for everyone. Cursor should have said upfront what Kimi was doing in this story. At the same time, if Composer 2 is close enough to more expensive models in real use, the question returns to the frontier labs: how much of their pricing reflects unique capability, and how much reflects trust in a closed brand?
