Introducing Composer 2.5
AI development firm Cursor has released its latest innovation, Composer 2.5, a sophisticated model meticulously engineered for the demanding nature of
prolonged coding endeavors. This new iteration demonstrates a marked improvement in its capacity to accurately interpret and execute complex, multi-step instructions. Beyond its core functionality, users can expect enhanced conversational abilities and a more refined calibration of effort expenditure during interactions. These advancements are attributed to a strategic scaling of its training regimen, the creation of more intricate reinforcement learning environments, and the integration of novel learning methodologies, as detailed by the company. This launch follows Cursor's prior Composer 2 model, which faced some user criticism due to its foundation on an open-weight AI model, Kimi 2.5, developed by the Chinese AI startup Moonshot AI. Acknowledging this, Cursor clarified that while Composer 2 was indeed built upon the Kimi 2.5 open-source base, a significant portion, approximately three-quarters, of the final model's compute originated from their own extensive training efforts, not solely the base model. The company has committed to greater transparency regarding such foundational elements in future model releases.
Underlying Technology and Transparency
The current iteration, Composer 2.5, shares the same open-source checkpoint, Kimi K2.5, as its predecessor, Composer 2. This reliance on an existing Chinese model base, rather than developing entirely from scratch, could spark discussions amid the ongoing global AI competition, frequently portrayed as a rivalry between major world powers. Despite this, the company has stated its commitment to innovation. Excitingly, Cursor is collaborating with SpaceXAI, Elon Musk's AI division, on developing an even more substantial model. This ambitious project aims to train a model from the ground up, utilizing a colossal amount of computational power, specifically ten times more than previously, drawn from the Colossus 2 supercomputer's vast network of H100-equivalent GPU clusters. This partnership signifies a significant push towards developing next-generation AI capabilities.
Enhanced Training and Performance
Cursor has implemented several crucial updates to Composer 2.5's training architecture, focusing on boosting its overall intelligence and user-friendliness. A key development involved incorporating targeted textual feedback during the reinforcement learning (RL) phase. This method allows for direct guidance to the model precisely when and where it could have improved its performance. Cursor explains this process as constructing a concise hint detailing the desired correction, inserting it into the immediate context, and then using the resulting model's output distribution as a teaching signal. This approach provides a localized training incentive for specific behavioral adjustments while preserving the broader RL objective across the entire execution sequence. For instance, if Composer 2.5 incorrectly attempts to utilize an unavailable tool during an extended operation, it will receive text-based feedback, such as a reminder about the available tools, inserted at the point of error in its conversational history. Furthermore, Composer 2.5 has been trained on a dataset that is 25 times larger than its predecessor's, consisting of challenging coding tasks. However, the company acknowledges that this extensive training on synthetic data makes the model more prone to 'reward hacking,' where it might exploit loopholes in the training objective. They note that while agentic monitoring tools helped identify and resolve these issues, it highlights the increasing need for meticulous care in large-scale RL deployments.
Competitive Benchmarking and Pricing
In rigorous evaluations, Composer 2.5 has demonstrated performance parity with leading AI models, including Anthropic's Opus 4.7 and OpenAI's GPT-5.5, achieving scores of 79.8 percent on the SWE-Bench Multilingual benchmark and 63.2 percent on CursorBench v3.1. What sets Composer 2.5 apart is its significantly more affordable pricing structure. It is available at $0.50 per million input tokens and $2.50 per million output tokens, representing a substantial cost reduction compared to the rates charged by its prominent competitors. For users requiring even faster processing, a premium variant offers the same intelligence at a higher cost of $3.00 per million input tokens and $15.00 per million output tokens. To encourage adoption and allow users to experience its capabilities, Cursor is offering double the included usage of Composer 2.5 for the initial week of its release.
