There is a battle to have the AI ​​model that programs best. And a good, pretty and very cheap rival has appeared in it: Cursor

Cursor has introduced Composer 2.5a generative AI model specifically intended for one thing: programming well. How good? Well, according to this startup, it does it as well as the best models of the moment, Claude Opus 4.7 and GPT 5.5, but it also does it for a lower cost. The challenge is striking not only because of what it means for Cursor, but because of how they have created that model: it turns out that it is based on a Chinese AI model.

AI models specialized in one thing. While OpenAI and Anthropic try to develop general-purpose models—they do a lot of things really well— Cursor you have decided to focus on a specific task. The AI ​​startup has created an AI model specialized in programming, and has done so by arguing that a billion parameters are not necessary to compete with the best. Devoting yourself to a single thing allows you to not only gain efficiency, but also costs. This is not a decathlete, but a specialist in the 200 m event, so to speak.

As good as GPT-5.5 or Claude Opus 4.7? That’s what they say in Cursor, because according to their tests with several specific programming benchmarks, the performance is on par with those two models that today are the great references both in programming and in other areas.

And much cheaper. These results are also especially interesting when we add the cost factor. The average cost per task in the CursorBench 3.1 benchmark showed that Composer 2.5 managed to solve almost 65% of all tests for a cost of just $0.3. Opus 4.7 max and GPT-5.5 xhigh managed to reach that 65%, but at much higher costs: just over 4 dollars in the case of GPT, and 11 dollars in the case of Opus. The difference is abysmal. He API access price demonstrates the differences: 0.5 dollars per million input tokens 2.5 dollars per million output tokens, when Claude Opus 4.7 is 5/25 and that of GPT-5.5 is 5/30 respectively.

Textual feedback. Unlike models that only learn from the final result, Composer 2.5 has been trained with a reinforcement learning technique (Reinforcement Learning) that allows us to offer clues about what is happening if errors are being made. This allows the model to recalibrate and act as a transparent teacher. One that also corrects word by word as it solves the exercise, not just when seeing the final result. 85% of the training budget has been dedicated exclusively to reinforcement learning, calibrating the model not for chat, but to execute code refactorings or fix bugs in real time.

A model “born” in China. Those responsible for Cursor themselves have explained that Composer 2.5—like its predecessor, Composer 2launched at the end of March—is a model derived from Kimi K2.5, the AI ​​model of the Chinese startup Moonshot. Although that is the basis, already in Composer 2 the training and post-training tasks manage to improve the behavior in a very notable way in programming benchmarks and also in others such as Terminal Bench that evaluate the agentic behavior of these models.

Cursor gets older. This startup became famous for creating a programming AI agent that was a pioneer in that fever we live for vibecoding. The user experience is no longer that of programming, as in traditional IDEs (Integrated Development Environments), but rather that of directing the machine to program it for you. Composer 2.5 doesn’t just program: it understands the structure and relationships between files, and turns Cursor into a much more competitive AI company, because it no longer depends on being able to work with Anthropic or OpenAI models, for example. Having both the AI ​​agent and the model processing everything makes it a much more competitive solution.

Elon Musk has Cursor in his sights. Cursor’s good performance has led to growing interest in buying this company even before it becomes too big. Elon Musk knows this well and Grok, xAI’s model, is not so popular in the programming field. In April we learned that SpaceX had reached an agreement that gives you the option to buy Cursor for 60,000 million dollars. It would be a promising deal for both, because Composer 2.5 has already used Colossus’ infrastructure to train, and xAI could thus try to gain market share in the juicy enterprise sector.

In Xataka | Elon Musk knows that TSMC is overwhelmed: Terafab is his idea to completely change the global chip industry