AI coding startup Cursor has launched a brand new mannequin referred to as Composer 2.5 that has been particularly skilled for long-running coding duties.
Composer 2.5 additionally follows advanced directions extra reliably, moreover different behavioural enhancements resembling communication fashion and energy calibration, Cursor mentioned in a weblog publish on Monday, Could 18. The enhancements in Composer 2.5 have come from scaling coaching, producing extra advanced RL environments, and introducing new studying strategies, as per the corporate.
Composer 2.5’s debut arrives months after Cursor’s Composer 2 mannequin that drew some backlash after customers discovered that the mannequin was a RL-modified model of Kimi 2.5, an open-weight AI mannequin lately launched by Moonshot AI, a Chinese language AI startup backed by Alibaba and HongShan (previously Sequoia China).
Acknowledging that Composer 2 was constructed on high of Kimi 2.5, Lee Robinson, Cursor’s vp of developer schooling, mentioned, “Yep, Composer 2 began from an open-source base!” “Solely ~1/4 of the compute spent on the ultimate mannequin got here from the bottom, the remaining is from our coaching,” he added.
“It was a miss to not point out the Kimi base in our weblog from the beginning. We’ll repair that for the subsequent mannequin,” Aman Sanger, the co-founder of Cursor, mentioned.
To make sure, the newest 2.5 variant can be constructed on the identical open-source checkpoint (Kimi K2.5) as Composer 2. In addition to not creating its coding mannequin from scratch, Cursor counting on a Chinese language mannequin base may doubtlessly elevate considerations amid the worldwide AI arms race that’s typically framed as an existential battle between the US and China.
Introducing Composer 2.5, our strongest mannequin but.
It’s extra clever, higher at sustained work on long-running duties, and extra dependable at following advanced directions.
For the subsequent week, we’re doubling the included utilization of the mannequin. pic.twitter.com/N87ojcXlOC
— Cursor (@cursor_ai) May 18, 2026
Final yr, the US-based startup raised a $2.3 billion spherical at a $29.3 billion valuation, and is reportedly exceeding $2 billion in annualised income. In April, Elon Musk-owned SpaceX, which can be now the father or mother agency of xAI, introduced plans to amass Cursor for $60 billion someday later this yr.
Cursor on Monday mentioned that it’s already working with SpaceXAI (the brand new AI division of SpaceX) to coach a “considerably bigger mannequin” from scratch utilizing 10 occasions extra whole compute from tens of millions of H100-equivalent GPU clusters that make up the Colossus 2 supercomputer.
Story continues beneath this advert
Beneath the hood
In the meantime, Cursor mentioned that it made a number of new modifications to the coaching stack of Composer 2.5 that targeted on bettering mannequin intelligence and usefulness. For starters, Composer 2.5 was skilled with focused textual suggestions throughout reinforcement studying (RL), which allowed them to supply suggestions on to the mannequin on the level within the trajectory the place the mannequin may have behaved higher.
“For a goal mannequin message, we assemble a brief trace describing the specified enchancment, insert that trace into the native context, and use the ensuing mannequin distribution as a trainer,” Cursor mentioned. “This offers us a localised coaching sign for the conduct we wish to change, whereas nonetheless retaining the broader RL goal over the complete trajectory,” it added.
For instance, when Composer 2.5 makes an attempt to name a software that isn’t out there throughout an extended rollout, it would obtain textual content suggestions on the error the place a touch resembling “Reminder: Accessible instruments…” is inserted within the context of the problematic flip.
Composer 2.5 can be skilled on 25 occasions extra artificial information (within the type of troublesome coding duties) than its predecessor. Nevertheless, Cursor warned that the newest mannequin is extra vulnerable to reward hacking as a consequence of coaching on artificial duties. “We have been capable of finding and diagnose these issues utilizing agentic monitoring instruments, however they display the rising care essential for big scale RL,” it mentioned.
Story continues beneath this advert
Efficiency on benchmarks
Composer 2.5 matched main AI fashions resembling Anthropic’s Opus 4.7 and OpenAI’s GPT-5.5 when evaluated on benchmark checks resembling SWE-Bench Multilingual (79.8 p.c) and CursorBench v3.1 (63.2 p.c).
Nevertheless, Composer 2.5 is less expensive to make use of per job as it’s priced at $0.50 per million enter tokens and $2.50 per million output tokens, a fraction of what Anthropic and OpenAI at present cost.
There’s additionally a sooner variant with the identical intelligence at $3.00 per million enter and $15.00 per million output tokens. Composer 2.5 contains double utilization for the primary week.