--- title: Cursor releases Composer 2, its third in-house coding... description: Cursor has released Composer 2, its third in-house coding model in five months. It scores 61.3 on CursorBench and 73.7 on SWE-bench Multilingual, significantly... canonical: https://daily.dev/posts/cursor-releases-composer-2-its-third-in-house-coding-model-in-five-months-mikxonmwg twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Cursor releases Composer 2, its third in-house coding model in five months | daily.dev og:description: Cursor has released Composer 2, its third in-house coding model in five months. It scores 61.3 on CursorBench and 73.7 on SWE-bench Multilingual, significantly... og:url: https://daily.dev/posts/cursor-releases-composer-2-its-third-in-house-coding-model-in-five-months-mikxonmwg og:image: https://api.daily.dev/og/posts/mIKxoNmwG.png og:image:alt: Cursor releases Composer 2, its third in-house coding model in five months og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Cursor releases Composer 2, its third in-house coding model in five months **[Collections](https://daily.dev/sources/collections)** · 3 min read · 1 upvotes · 0 comments ## Summary Cursor has released Composer 2, its third in-house coding model in five months. It scores 61.3 on CursorBench and 73.7 on SWE-bench Multilingual, significantly outperforming Composer 1.5. The headline is pricing: $0.50/$2.50 per million tokens — an 86% reduction from Composer 1.5 and roughly 10x cheaper than Claude Opus 4.6. Technically, it introduces continued pre-training before RL (built on Kimi 2.5) and a novel 'self-summarization' technique that compresses context to ~1,000 tokens mid-task, reducing compaction errors by 50% and enabling hundreds of sequential actions without losing coherence. Early testing shows ~200 tokens/second and solid performance on iterative coding tasks. Cursor's team of ~40 researchers is entirely focused on coding, positioning the company between pure app maker and model provider. ## Content Cursor has released Composer 2, its latest in-house coding model and its third in five months. It's now available directly in the Cursor editor. ## Benchmarks and pricing Composer 2 scores 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0, and 73.7 on SWE-bench Multilingual. For comparison, Composer 1.5 scored 44.2, 47.9, and 65.9 on those same benchmarks. On Terminal-Bench 2.0, it edges out Claude Opus 4.6 (58.0%) and sits close to GPT-5.4. The pricing is the headline: $0.50 per million input tokens and $2.50 per million output tokens. That's an 86% reduction from Composer 1.5 and roughly 10x cheaper than Claude Opus 4.6 ($5/$25 per million tokens). A faster variant is available at $1.50/$7.50 per million tokens, and Cursor has made the faster version the default. ## What's technically new Two things stand out under the hood. First, Composer 2 is the first version to use continued pre-training before reinforcement learning, giving it a stronger base before the RL work begins. The base model is Kimi 2.5. Second, and more interesting, is what Cursor calls "self-summarization" - a compaction-in-the-loop RL technique. When the model hits token-length limits mid-task, it compresses its own context window down to roughly 1,000 tokens rather than relying on a sliding window or external text summary. The model is trained to do this compression itself, which Cursor says reduces compaction errors by 50% compared to traditional approaches. The practical effect is that the model can handle tasks requiring hundreds of sequential actions without losing the thread. ## Real-world performance Early hands-on testing has been broadly positive. Developers report around 200 tokens per second, and the model handles iterative vibe coding tasks well - building things like 3D simulators and tile-based games with reasonable coherence across many prompts. That said, benchmarks aren't the whole story. Cursor's own team notes that Opus 4.6 still has a better writing style for non-coding tasks, and that benchmark results shouldn't be treated as universal rankings. The team is actively dogfooding Composer internally and flagging quirks - things like bad markdown formatting or overly verbose summaries - to fix through RL penalties. ## The bigger picture Composer 2 marks one year since Cursor started its large model training efforts. The team is now around 40 researchers, entirely focused on coding. No email assistants, no tax tools, no chatbots - every training run is aimed at software engineering. Cursor's position is deliberately in between pure app maker and model provider. The goal is to combine the best API models with domain-specific models built in-house, optimizing the model, interface, and token economics together as a single system. Whether that focused approach can keep pace with labs spending orders of magnitude more is the interesting question - but Composer 2 is at least a credible data point that it can. --- Tags: [#llm](https://daily.dev/tags/llm), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning), [#ai-assisted-development](https://daily.dev/tags/ai-assisted-development) [View this post on daily.dev](https://daily.dev/posts/cursor-releases-composer-2-its-third-in-house-coding-model-in-five-months-mikxonmwg) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Cursor releases Composer 2, its third in-house coding model in five months","url":"https://daily.dev/posts/cursor-releases-composer-2-its-third-in-house-coding-model-in-five-months-mikxonmwg","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/cursor-releases-composer-2-its-third-in-house-coding-model-in-five-months-mikxonmwg"},"datePublished":"2026-03-19T16:35:31.704Z","dateModified":"2026-03-24T16:08:30.309Z","description":"Cursor has released Composer 2, its third in-house coding model in five months. It scores 61.3 on CursorBench and 73.7 on SWE-bench Multilingual, significantly...","image":"https://pbs.twimg.com/media/HDyS34HbAAAEOoq.png","thumbnailUrl":"https://pbs.twimg.com/media/HDyS34HbAAAEOoq.png","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/cursor-releases-composer-2-its-third-in-house-coding-model-in-five-months-mikxonmwg","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,reinforcement-learning,ai-assisted-development","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Cursor releases Composer 2, its third in-house coding model in five months"}]} ```