What shipped on May 18
Composer 2.5 is Anysphere's in-house coding model, built on the same base as its predecessor: Moonshot AI's open-weight Kimi K2.5 checkpoint, continued pretraining and reinforcement-learning on long-horizon coding tasks. It is priced at $0.50 per million input tokens and $2.50 per million output tokens, with a faster variant at $3.00 and $15.00. Cursor describes it as trained on 25 times more synthetic tasks than Composer 2, using a technique it calls feature deletion, where the model has to reimplement code that was deliberately removed and its work is checked against the project's test suite. [1]
The announcement also describes infrastructure changes behind the training run: a distributed optimizer it calls Sharded Muon, paired with what it terms dual-mesh hybrid sharded data parallelism, reaching a 0.2-second optimizer step on a model with a trillion parameters. It says a larger model trained from scratch is planned, using ten times the compute on SpaceX's Colossus 2 cluster. [1]
A number that only Cursor has measured
Two months earlier, Cursor's own post for Composer 2 reported 73.7 on SWE-bench Multilingual, a coding-agent benchmark that grades whether a model can resolve real GitHub issues across several programming languages. The summary above puts Composer 2.5 at 79.8 on the same benchmark, alongside comparison figures for Claude Opus 4.7 and GPT-5.5 that Cursor presents in chart form rather than in the post's text. [1][2]
Atlas interpretation: The improvement from Composer 2 to Composer 2.5 on SWE-bench Multilingual is consistent between the two Cursor posts, so the trend is Anysphere's own reporting held constant across two releases rather than a single cherry-picked figure. The comparison to Opus 4.7 and GPT-5.5 is a different kind of claim: it is Cursor's chart, run on Cursor's harness, with no independent lab result available to check it against. "Matches the frontier" is a vendor's characterization of its own scoreboard. [1][2]
A release from a company already under an acquisition option
Composer 2.5 shipped about a month after SpaceX's option deal, in which SpaceX paid for the right to either license Cursor's models for roughly $10 billion or buy Anysphere outright for $60 billion later in 2026. SpaceX exercised the acquisition path on June 16, four weeks after this release, and the deal closed that August. [3]
Atlas interpretation: That timing puts Composer 2.5 in an odd position: a coding model marketed on price and independence, shipped by a company that had already signed away its own future to the compute supplier it was competing to be cheap for. The "tenth the price" claim in the event's title reads differently once the $60 billion acquisition is in view: within a year, Anysphere would be reporting to the company that also owns the GPU cluster its next model is meant to train on. [3]
Sources
- Introducing Composer 2.5
Cursor · May 18, 2026
- Introducing Composer 2
Cursor · Mar 19, 2026
- SpaceX announces $60 billion Cursor deal to boost AI coding
Yahoo Finance · Jun 16, 2026