Cursor launches Composer 2.5 model

Cursor released Composer 2.5 on May 18, adding a new in-house coding model aimed at long-running software tasks and lower token costs. The company said the model improves on Composer 2 in sustained work, instruction following and collaboration behavior, while keeping its standard price at $0.50 per million input tokens and $2.50 per million output tokens. A faster default tier is priced at $3 per million input tokens and $15 per million output tokens, according to Cursor’s changelog. The launch matters because Cursor is no longer presenting itself only as an interface on top of outside frontier models. Indian Express reported on May 19 that Cursor also said it is working with Elon Musk-owned SpaceXAI to train a significantly larger model from scratch using 10 times more total compute from the Colossus 2 supercomputer. Cursor’s own materials for Composer 2 and 2.5 emphasize internal model training, reinforcement learning on long-horizon coding tasks and pricing aimed at developer usage rather than one-off demos. (cursor.com) ### What actually changed in Composer 2.5? Cursor’s May 18 changelog described Composer 2.5 as “a substantial improvement in intelligence and behavior over Composer 2.” The company said the model is better at sustained work on long-running tasks, follows complex instructions more reliably and is more pleasant to collaborate with. Cursor did not publish a fresh benchmark table in the changelog entry it made public, but it did position the release as a behavioral and workflow upgrade rather than a full reset of the product line. (indianexpress.com) Kingy AI, in a May 19 review of the release, said the most defensible reading is incremental gains on raw capability and more meaningful gains on harder-to-measure areas such as tool-call reliability and instruction following. That framing matches Cursor’s own emphasis in the release note. ### Why is pricing central to this launch? (cursor.com) Cursor kept the standard Composer 2.5 price at $0.50 per million input tokens and $2.50 per million output tokens, while raising the price of the faster default tier to $3 and $15 respectively. By contrast, Composer 2 launched in March with the same standard pricing but a faster variant at $1.50 per million input tokens and $7.50 per million output tokens. (kingy.ai) That means the company is separating low-cost long-horizon work from higher-throughput usage more explicitly than before. TestingCatalog reported the new model is marketed as up to 10 times more efficient than similarly capable models. Cursor’s own public changelog does not use that exact phrasing, but its pricing and positioning point in the same direction: lower-cost sustained coding work is the selling point. (cursor.com) ### Why does the SpaceXAI detail stand out? Indian Express reported that Cursor said it is training a much larger model with SpaceXAI using Colossus 2 compute. If that effort progresses, Cursor would be moving further from the role of IDE wrapper and deeper into model ownership and infrastructure partnerships. Kingy AI said the release suggests Cursor is building identity as both a model company and an agent company, not just a client for OpenAI or Anthropic systems. (testingcatalog.com) That distinction matters for enterprise buyers because coding assistants are increasingly bought on three layers at once: the base model, the workflow product and the price structure. Cursor’s own product materials already tie models to cloud agents, development environments, usage pools and team controls, showing how the company is packaging model performance with operational tooling. (indianexpress.com) ### Where does this leave the coding-assistant market? Cursor’s March and May releases show a faster cadence of in-house model updates inside a broader coding platform. Composer 2 was introduced on March 19 with benchmark claims including 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0 and 73.7 on SWE-bench Multilingual. Composer 2.5 arrived less than two months later with the company stressing long-running task behavior and reliability instead of headline benchmark jumps. (cursor.com) Cursor said Composer 2.5 includes double usage for the first week, and the company directed users to its model docs and announcement for more detail. The next concrete signal will be whether Cursor publishes fuller benchmark data for 2.5 or gives more specifics on the larger model it says it is training with SpaceXAI. (cursor.com 1) (cursor.com 2)